The Bowl and the Black Box.

Nobody asks a search box what to feed their dog

Ask a pet owner what they want to know about their dog's food and you will not get a keyword. You will get a paragraph.

She's seven now, she's a working cocker, she's been bringing her breakfast back up about twice a week since we moved, the vet says nothing's wrong, I'd rather not spend forty pounds a bag but I would if it helped.

Nobody has ever typed that into a search box. Until recently there was nowhere to put it. You compressed it — “sensitive stomach dog food” — and accepted a page of results that answered a fraction of what you asked. Or you asked someone. The vet, the breeder, the woman in the pet shop who has been there eleven years, the forum where four hundred people have owned a working cocker with a delicate stomach.

That compression has stopped. The question can now be asked whole, and something answers it whole.

So we asked it. In August 2026, from a UK connection, we put thirty real UK pet-owner questions to ChatGPT, Gemini, Perplexity, Google's AI Overviews and Amazon.co.uk's Rufus — 234 captures in total, every answer dated and screenshotted, every citation logged.

We expected the machines to dodge the brand question and answer in generalities. They did not. Asked what's the best food for a dog with a sensitive stomach, the surfaces named brands in three quarters of the answers that appeared. They were confident, they were specific, and — this is the part that should concern anyone in this category — they largely named the same four.

The rest of this piece is about what that means for anyone who writes, owns, or is judged on a pet food product page in the United Kingdom.

1. The category where the buyer is never the consumer

Start with the thing that makes pet nutrition strange, because everything else follows from it.

In almost every category you can name, the person paying eventually finds out whether they were right. You eat the meal. You wear the coat. The feedback loop closes, sometimes brutally, and next time you buy differently.

Pet nutrition does not work like this. The buyer is never the consumer. The dog cannot say the food disagreed with it. The cat cannot explain that it is bored of the flavor, or that the kibble is too large, or that the thing making it uncomfortable started three weeks ago and is not the food at all. The owner observes a shine on a coat, a firmness in a stool, an enthusiasm at a bowl — proxies, all of them, noisy and slow and open to interpretation.

There is one close cousin: infant nutrition. Same structure, same stakes, same silence at the other end. And the published work on infant formula shows what buyers do when the feedback loop is broken. They substitute authority for evidence. A 2025 study in Maternal & Child Nutrition found that 58% of infant formula product labels evoked medical authority — health-professional endorsement deployed specifically to shore up claims the buyer had no way to test. Research on formula purchase intention published in Foods found intent driven indirectly, through perceived quality and governance signals, rather than through anything resembling direct experience.

That transfer from infant formula to pet food is an inference rather than a measured finding, and should be read as one. But it is well-founded, and it explains three things about this category that are otherwise puzzling.

First, the category is unusually receptive to an authoritative-sounding answer. It always has been. The vet, the breeder, the pet shop owner, the forum — these are not information sources of last resort, they are the primary mechanism by which the decision gets made. An assistant that speaks with total confidence and no visible uncertainty is not a novelty here. It is a new occupant of a role the category has kept open for a century.

Second, content is not decoration in this category. It is the substitute for an experience the buyer can never have. Thin content in a category with a working feedback loop is a missed opportunity. Thin content here is a void where the entire basis of the decision was supposed to sit.

Third, trust signals outrank persuasion. Credentials, provenance, substantiation, and other owners' experience do the work that taste and preference do elsewhere. This is not a stylistic preference. It is what happens when nobody can verify anything directly.

Humanization raised the stakes

The pet became family, in a sense strong enough to change purchasing. The exact number depends entirely on the question asked, and it is worth being precise, because the figures circulating in the trade are not interchangeable. Dogs Trust's National Dog Survey 2024, with over 406,000 participants, found 99% of owners describe their dog as part of the family and 89% call them their best friend. YouGov, asking the stricter question — whether pets are family members on the same level as a human — finds 59% of dog owners agreeing. Both are UK-specific and both are real; they are not measuring the same thing.

The market moved with it. UK Pet Food's population research, conducted by Kantar with fieldwork in January 2026 across roughly 8,951 interviews, puts 36.5 million pets in around 18 million homes — 62% of UK households. 15.5 million dogs. 13.1 million cats. The UK pet food market was worth £4.3 billion in 2025, up 4.9% year on year.

Growth is concentrating at the credentialed end. Ken Research puts the organic segment at approximately £280 million with double-digit growth, and projects premium approaching £1 billion. Mintel documents functional formats — digestive, joint, skin and coat, hydration — growing faster than the category.

The distinction matters more than it looks. Humanization is not simply pushing owners toward higher prices. It is pushing them toward credentials — toward foods that can prove something. Hold that thought, because the probe found that the machines have already made exactly the same move, and made it much more aggressively than owners have.

2. The advisor is under strain

If the category outsources judgment to authority, the obvious question is what happens when the authority weakens.

We have an unusually well-documented answer, because the Competition and Markets Authority spent nearly three years examining it.

What the CMA found

In 2013, large veterinary groups owned around 10% of UK local practices. By the time of the investigation that share had passed 60%, and the CMA's final report found average veterinary prices rose 63% between 2016 and 2023, well above inflation — with consumer detriment estimated at roughly £1 billion over five years. Fewer than half of confirmed large-group customers, 47%, knew their practice was part of a large group at all.

A channel where prices rose sharply, ownership became opaque, and the customer often could not find out what anything cost before walking in.

The CMA published that report on 24 March 2026 and must make a legally binding Order by 23 September 2026, bringing mandatory price lists, a prescription-price remedy, a price-comparison facility and compulsory disclosure of large-group ownership. Whether the remedies work is not the point. The point is that the category's authority is in the middle of a public, adversarial correction — during which owners have been told, in a government report, that they have probably been overpaying.

The vet is still the nutrition authority — and is not always available for it

Veterinary professionals remain the category's nutrition authority. A PLOS One study with fieldwork between October 2023 and January 2024 found UK and Irish veterinary and veterinary-nursing students regard veterinarians and vet nurses as the leading nutrition authorities. A Canadian survey of 2,181 owners in the Canadian Veterinary Journal found the veterinary team was the primary nutrition source for 43.6%, with the internet already primary for 24.6%.

But the advice is not consistently delivered. Research from the Purina Institute — brand-commissioned, which we flag rather than launder, and US in origin — found that while 96% of owners trust their veterinary team for nutrition advice, only 22% of veterinary team members proactively raise nutrition at every appointment.

The authority is trusted, getting more expensive, under a transparency correction, and does not consistently offer advice on the specific question this article is about.

And the alternative arrived

General AI use in the UK went from marginal to mainstream in under two years. Ofcom's Online Nation 2025, published 10 December 2025, recorded 1.8 billion UK visits to ChatGPT in the first eight months of 2025 against 368 million a year earlier — roughly five-fold. Ofcom's Adults' Media Use and Attitudes found 54% of UK adults using AI tools in late 2025, up from 31%. Around 30% of Google searches now return an AI overview.

There is a real generational skew — 79% of 16-to-24s against roughly 18% of over-65s — and honesty requires naming it, because the older, higher-spending pet owner is the least likely to have made the switch.

What the machine actually does with a vet question

Here the probe has something specific to add, and it is not what we expected.

Across our vet-displacement block, the surfaces deferred to a vet in 33% of answered runs. Two thirds of the time, the machine simply answered.

But deferral is the smaller finding. The larger one is what it says when it does not defer. In the vet-displacement block, 77% of all brand mentions were clinically positioned or veterinary-channel brands. Asked whether to change the food or see the vet, the machine frequently does both jobs at once: it gives general advice and it names the brands the veterinary channel has always recommended.

On the sharpest prompt in the set — is my vet recommending a food because it's best or because they sell it? — the surfaces did not dodge. They acknowledged the conflict, explained the commercial structure of prescription diets, and in several runs supplied the owner with a script of questions to put to their vet. One Google AI Overview answered the question almost entirely by surfacing forum threads on the subject, including the recurring internet argument that vets recommend a specific handful of brands because those brands meet WSAVA guidelines that most manufacturers do not.

That is worth sitting with. The machine is not replacing the vet's authority. It is repeating the vet's recommendation set, at scale, without the consultation, and while explaining to the owner why the vet might be conflicted.

3. The product page has three readers — and the answer set collapses

For twenty years, a product page had one reader with one set of habits. A person scanned it, inferred generously, and filled in whatever was missing from the picture on the pack.

That page now has three readers, and the other two are nothing like the first.

Reader one, the human, is still there, still scanning, and crucially infers freely. Shown a photograph of a grown dog, the human concludes the food is for adults. Shown “complete nutrition,” the human concludes this is a meal. None of those inferences is stated anywhere on the page. All are made anyway. What the human rewards is persuasion. What is at stake is a sale.

Reader two, the assistant — ChatGPT, Gemini, Perplexity, and on the retailer's turf, Rufus — is human-guided but reads and reasons over the listing on the owner's behalf. It extracts. It infers poorly. It cannot reliably deduce “complete” from “nutrition,” or “adult” from a photograph, or “grain free” from an absence. The scale is not speculative: Amazon's Q4 2025 earnings, reported February 2026, disclosed Rufus was used by more than 300 million customers in 2025 and drove nearly $12 billion in incremental annualized sales. On the Q3 2025 call, Andy Jassy said monthly active users were up around 140% year on year and that shoppers engaging with Rufus were roughly 60% more likely to complete a purchase. Rufus is live on Amazon.co.uk. What is at stake is inclusion in the consideration set — or silent omission from it, which produces no signal at all.

Reader three, the agent, acts. In November 2025 Amazon added account memory, automatic cart additions, price alerts and target-price auto-buy to Rufus, and Jassy described it as “the AI agent.” It matches against constraints, reads once, decides, and stops reading. What is at stake is not one order. It is every order after it.

What the probe found — and what it killed

We built this study to test a specific claim: that AI answers nutrition questions with functional attributes rather than brand names, so brand equity fails to transfer into the AI layer.

That claim is wrong, and we are retiring it.

Across 106 runs where a surface produced an answer, 56 answers were brand-led — the recommendation itself was one or more named brands — against 26 led by functional attributes. Counting more loosely, a brand was named somewhere in 63 of the answered runs; we lead with the stricter figure throughout. At the functional need-state layer — where the thesis specifically predicted brands would disappear — 67% of answers named brands. On our headline pair, the contrast the study was designed around barely existed: asked what's the best food for a dog with a sensitive stomach, six of eight answers named brands; asked which dog food brands are most trusted in the UK, six of seven did.

Brand equity transfers into the AI layer perfectly well. That is not the problem.

The problem is that the answer set collapses.

Ask the broad trust question and the machine gives a wide, distinctly British answer: 29 different brands across the runs, with the top three accounting for just 26% of mentions. Specialist names, premium challengers, heritage brands, own-label — the full shape of the UK category.

Ask a functional need-state question and the field closes. Twelve brands, with the top three taking 63% of all mentions. Eighty-six percent of every brand mention at the need-state layer went to clinically positioned or veterinary-channel brands — a group drawn from the three largest manufacturers in the category, including the co-author's.

That is the finding. Not that brands vanish, but that at the exact moment the owner asks the question that matters most — the compound, need-state, high-anxiety question — the machine stops surveying the category and starts reciting a shortlist. A brand that appears in the broad answer and is absent from the narrow one has not lost visibility. It has lost the only visibility that converts.

And the shortlist is unstable. Across identical Tier 1 prompts repeated on the same surface, more than half of prompt-surface cells — 16 of 30 — returned a different set of brands between runs, and in a third of cells the brand named first changed, including cells where one run named a brand and another named none at all. AI visibility is not a ranking that can be held. It is a probability that can be shifted.

Your listing is not the only thing the assistant reads

We logged 1,785 citations. The distribution is the most uncomfortable table in this study.

Manufacturer-owned content — every brand site, every brand blog, every piece of owned content from every manufacturer in the sample combined — accounted for less than six percent of what the machines cited. Reddit alone, at 179 citations, was cited more often than all brand-owned content put together.

Across 297 distinct domains, the most-cited single source was Amazon.co.uk, followed by Reddit and then Which?. Review sites and forums together made up 54% of everything cited.

The category's authority has migrated, and it did not migrate to brands.

Having the credential is not the same as stating it

Two further findings from Genrise's own research — first-party, and flagged as such for the same reason we flag the vendor-commissioned figures elsewhere in this piece — explain what “infers poorly” costs in practice.

The first is counterintuitive enough to be worth stating carefully. In a large study of retail-AI recommendation behavior across 610,096 events and 22,824 products, the category best-sellers the assistant never once recommended carried more customer reviews — a median of 1,114 — than the products it recommended occasionally, at 651, on the same star rating. Reviews, ratings, a recognized name: all present, and the assistant passed over them anyway. The separating variable was whether the page answered the shopper's underlying question in words. Products the assistant recommended most often did so 88.8% of the time. The never-recommended group managed 23.6%. This study spans eight categories across three markets and is not pet-specific, so it should be read as a directional cross-category parallel rather than a pet-nutrition fact.

The second is pet-specific, and it identifies the mechanism precisely. Testing wet dog food pages against shopper questions — 878 US and 5,762 UK question-by-page judgments, over 122,000 individual phrase checks — pages stating only the single most obvious keyword of a need-state claim were credited with answering the question on well under half of the pages judged. Pages stating the same claim with its full surrounding vocabulary were credited on a clear majority. Not a stronger claim. Not a different claim. The same permitted claim, written out properly. The gap ran 20 to 40-plus percentage points across every claim type tested, widest of all on problem-framing language — which is exactly the register an anxious owner writes in.

That is the whole difficulty of reader two in one sentence. It is not withholding credit from brands that lack credentials. It is withholding credit from brands that have them and did not spell them out.

Retailer AI is not open-web AI

This is the commonest error in the market, and the probe settles it cleanly.

Every one of Rufus's 210 logged citations resolved to a retailer or marketplace source. One hundred percent. Rufus reads Amazon and cites Amazon: a closed corpus. Meanwhile the open-web surfaces drew from 297 domains, over half of them reviews and forums.

The two surfaces also disagree about who exists. Own-label was named in 58% of brand mentions when open-web assistants were asked about supermarket brands — but appeared in just one of 54 Rufus runs. Rufus was also more brand-forward overall, returning brand-led answers in 78% of runs against 53% on the open web.

These surfaces read different corpora, cite differently, and reward different things. A strategy treating “AI visibility” as one surface is optimizing for the average of two things that do not average.

4. Why pet nutrition breaks first

Plenty of categories are exposed to AI-mediated discovery. Pet nutrition is exposed worse, for a structural reason worth stating precisely: the query here is almost never simple.

Consider what must be true before a recommendation is any good. Species. Life stage. Breed size. Body condition. A health consideration, or several. Format. Palatability history, which here means the last four things the animal refused. Feeding routine. Budget, expressed not as price per bag but price per day. And increasingly a sourcing preference held on principle.

Ten constraints, routinely, in a category where the average basket is under thirty pounds. Most categories do not compound like this.

Compound queries expose thin content faster than anything else, because thin content fails at the first unmet constraint rather than degrading gracefully. And Section 3 showed what that failure looks like in practice: not a lower ranking, but exclusion from a shortlist of four.

The probe also showed which attributes the machines actually reach for. Across all answered runs, the functional vocabulary was narrow and repetitive — sensitive stomach, wet, dry, natural, urinary, grain-free, kibble, then a long tail of digestibility, hydrolyzed protein, hypoallergenic, prebiotics and probiotics. That is a small, learnable set. It is also, almost exactly, the vocabulary of the clinical brands that dominate the shortlist.

The need-states with the most demand are the ones you cannot claim

Here is the trap, and it links directly to Section 8. The highest-demand need-states — digestive, urinary, joint, sensitive stomach, immunity — sit hard against the boundary of medicinal classification under UK law. The owner is searching for the sentence the brand is not permitted to write. We come back to exactly where that line falls.

Inflection points, and why the shelf reopens

It would be easy to read the agent as reading once and never returning. That is too pessimistic, and correcting it makes the argument both more honest and more useful.

Pets change. A puppy becomes an adult, then a senior. A cat is diagnosed. The family travels. A second animal arrives. Each is an inflection point — a moment the owner re-enters the category and the consideration set reopens. The question shifts from “what should I feed my dog” to “what should I feed my dog now that he is seven and his stomach is unreliable.” Different question, different constraints, and the winner of the first has no automatic claim on the second.

The probe found something quietly encouraging here. Life-stage questions — when do I move him onto adult food, does he need senior food yet, how do I change his food without upsetting his stomach — produced almost no brand mentions at all. They were answered as advice, and deferred to a vet in two of three answered runs. This is the one part of the category the shortlist has not captured. It is open ground.

We could not locate UK panel data on switching rates or trigger frequency, and we have not estimated it. The mechanism is well-founded; the rate is not sourced. One inflection point is well documented: UK Pet Food's guidance on transitioning gradually between foods exists because abrupt changes cause gastrointestinal upset — which in a retail context becomes a return and a one-star review blaming the food.

5. The eleven things your listing must be able to say

This is the working heart of the piece, and it is deliberately not a checklist of fields. Fields are how a PIM thinks. This is a list of questions a listing must be able to answer, each ending in something briefable.

A note on evidence first. The best-known content-to-conversion figures in this space are US in origin and vendor- or platform-commissioned, including Amazon's own. We use them as directional proxies and label them as such. We could not locate any independent, published UK pet-specific content-performance research. The page-level findings cited below are our own, and we flag them as first-party throughout rather than presenting them as neutral. The gap is worth naming either way: a category is being asked to invest in content on the strength of evidence that is either borrowed from another market and another aisle, or produced by the people recommending the investment.

1. Is this for my animal, right now?

Species, life stage, breed size, body condition — in words rather than implied by a photograph. The human infers “adult” from a picture of a grown dog. Neither of the other two readers can.

2. Is this a meal, or is it an extra?

The single most important nutritional signal in the category, and the one most likely to be missing or unusable. The probe makes the scale of the omission concrete: across 106 answered runs, “complete and balanced” surfaced as a named attribute eight times, and the word “complementary” appeared once in the entire captured corpus. The distinction that determines whether a food is a diet or a garnish is essentially absent from the machine's vocabulary — because it is largely absent from the listings it reads.

3. What is actually in it?

UK law permits category naming — “meat and animal derivatives,” “cereals.” It is entirely legal. It is also close to useless to a machine trying to determine whether a food contains chicken.

4. How much of what?

Analytical constituents, and the arithmetic an owner actually performs — including that a wet food's percentages are not comparable to a dry food's without adjusting for moisture. Almost no listing helps with this.

5. How much do I feed, and how do I switch?

Feeding guidance by weight, and transition guidance — the cheapest possible insurance against a return and a bad review, and missing more often than not.

6. What need does this meet — within the line?

The functional claim, on the permitted side of the medicinal boundary set by the Veterinary Medicines Regulations 2013. Supporting or maintaining normal function in a healthy animal is permitted; treating, preventing, or naming a condition is not.

7. What is not in it?

Free-from and exclusion claims, substantiated. Among the most searched attributes in the category — grain-free appeared 12 times in the probe's attribute set — and among the most exposed under the CAP Code, which applies to retailer product pages.

8. Where does it come from, and who made it?

Provenance and manufacture. In a proxy-buyer category these are trust signals doing load-bearing work, not brand color.

9. Who says it is good?

Credentials — the most regulated sentence on the page. “Vet recommended” is commercially powerful precisely because it borrows the authority Section 2 is about. Under the CAP Code it requires substantiation, and the standard for an endorsement claim of that form is exact and evidential.

And Section 3 explains why this item has become the most valuable on the list. The machines are not choosing brands on the strength of copy. They are choosing on the strength of clinical credentials the internet has already ratified.

10. What do other owners say — in a form that can be quoted?

Reviews and Q&A are not a reputation asset any more. They are a corpus — and per Section 3, reviews and forums are 54% of everything the machines cite.

11. What does it cost to feed?

Not price per bag. Price per day, per animal, at the stated feeding rate. PDSA's PAW Report 2025 explains why: 51% of owners worry about the cost of care, 86% say costs have risen, and two-fifths say ownership costs more than expected.

The surfaces will not all carry this

These eleven things cannot be executed identically everywhere. Amazon.co.uk supports title, bullets, A+ and Premium A+ content, brand store, video and Q&A — a wide canvas, publicly documented. Field structures, character limits, image counts and enhanced-content availability at Tesco, Sainsbury's, Asda, Ocado, Pets at Home and Zooplus are not published in any consolidated public form. We flag that as an evidence gap rather than fill it with an estimate. The practical consequence is that any content brief must be validated against each retailer's actual specification before it is written. Recommending a module a retailer does not support is worse than recommending nothing.

6. The subscription trap — and the stale shelf

Pet food is one of the most subscription-native categories in retail, and the reason is almost embarrassingly simple. The pet does not change. The recipe does not change. The bag runs out on a schedule that can be calculated.

That is the definition of a category ready for agentic commerce.

The size of it, honestly

A figure of roughly one third of UK online pet spend running through subscription circulates in the trade. We could not source it and we are not repeating it.

What we can publish is a retailer disclosure. Pets at Home Group plc reported subscription revenue at 15.2% of consumer revenue in FY2026, the year ended 26 March 2026, up from 13% the prior year. That is a share of total consumer revenue — retail and veterinary combined — not of online spend, so it cannot be converted into the one-third claim by rearrangement. It is a floor, from a named source, on a different denominator.

Alongside sits a subscription-first D2C cohort competing on a model where the first order is the only order involving a decision. Mintel recorded Butternut Box as the largest UK pet food advertiser by adspend in 2024, which tells you what the first order is worth when you keep the customer.

The shelf goes dark

An agent reads the shelf once — at the moment the subscription is set — and then the shelf goes dark.

Every subsequent order is invisible to competitors. No impression to win, no result to rank in, no moment of reconsideration to interrupt. The competitor is not losing the sale; the competitor is not present at the sale, because there is no sale in any sense a marketer would recognize. There is a repeat instruction being executed.

In a subscription category, the agent does not choose a product. It chooses a relationship, and defends it by default.

The probe's subscription block is small — two answered runs — and we will not over-read it. But both named brands, and one surfaced the retailer's own subscription programme alongside a clinical brand. There is no evidence here that the subscription-setting moment is more open than any other. If anything it looked narrower.

One methodological note that matters commercially: every Rufus capture in this study came from a fresh account with no pet purchase history. Rufus gained account memory in November 2025, and personalization to prior purchases is precisely the mechanism that would harden the shelf shut. We measured the open door. We did not measure what happens after it closes.

And it goes dark to reality

The agent defends a relationship with a product it may no longer accurately understand. Recipes get reformulated. Pack sizes change. Lines get delisted. An agent re-ordering nine months after the subscription was set is acting on what it learned once, against a catalog that has moved underneath it.

Be careful here: this is a mechanism, not a measurement. We could not find UK data on how frequently pet foods are reformulated, repackaged, resized or delisted. No catalog-churn dataset exists in the public record. Catalog change happens at scale and is disclosed — Pets at Home reported completing category resets across its dog and cat food ranges in FY26 — but that is evidence of change, not a frequency. The staleness argument is sound on logic. The rate at which it produces error is not sourced, and we decline to imply one.

The counterweight: the shelf reopens

The trap is real. It is not permanent. Section 4 set out the inflection points, and the probe found them to be the least brand-captured territory in the whole study — life-stage questions produced almost no brand mentions and deferred to a vet more often than any other block.

That changes what content is for. It is not only there to win the first order. It is there to be findable at re-entry, when the question has changed shape entirely — and, on this evidence, at re-entry the shortlist has not yet formed.

7. Private label is not the cheap option any more

There is a version of this section that reads as an attack on retailers. It would be both wrong and useless. What follows is market structure.

The size and the divergence

Mintel's UK pet food work puts private label at around half of dog-meal volume. Circana's European data puts private label at roughly 34% of value in pet food, against about 38% across European FMCG generally.

That volume-value gap is not an inconsistency to be resolved. It is the finding. Own-label moves a great deal of product at a lower average price — which is exactly what has been changing.

A figure that circulates and should not be borrowed: Kantar's own-label share of around 51% is all UK grocery, not pet food. Different denominator.

The cost-of-living period accelerated the shift. Circana documented European retailers growing private-label pet food share through 2022, with UK pet food value sales up 12.7% that year. PDSA's PAW data shows what owners did: around 9% swapped to a cheaper brand of pet food, while roughly 18% — about four million owners — cut back on their own weekly food shop to keep providing for their pets.

Read those together, because it is the most revealing pair in this section. The dominant response to cost pressure was not to feed the animal worse. It was to trade down on price while holding the nutrition standard, or to absorb the cost personally. That is a demand for credentials at a lower price point — precisely the demand own-label moved to meet.

The premiumization is real and documented

The premium own-label tier is a substantial competitive presence: Tesco Finest, Sainsbury's Taste the Difference, Waitrose and M&S own ranges, and in the specialist channel Pets at Home's AVA and Wainwright's lines. Nestlé Purina's acquisition of Lily's Kitchen showed the same movement from the other direction. Underneath sits a UK contract-manufacturing base — GA Pet Food Partners, Naturediet, Fold Hill among them — supplying retailers with formulations carrying human-grade language, transparent sourcing and functional positioning, produced to BRC and FEDIAF standards.

What the probe found — own-label has arrived, but only when asked about directly

This is the sharpest own-label finding in the study, and it cuts both ways.

Ask the open-web assistants directly about supermarket brands and own-label does extremely well: 58% of all brand mentions in that block went to own-label names — Tesco, Aldi, Sainsbury's, Lidl, Waitrose, and the specialist retailer's own lines — against 19% for clinical brands. Asked whether own brand is as good as the big names, the machines frequently said yes, and cited Which? while doing it. Which? was the third most-cited domain in the entire study.

Now ask a functional need-state question. Own-label's share of brand mentions is zero. Not low. Zero, across every sensitive-stomach, urinary, skin-and-coat, weight, joint and hydration run in the study.

And on Amazon.co.uk, own-label appeared in one of 54 Rufus runs — an artifact partly of what Amazon ranges, but a real commercial fact for anyone assuming own-label parity travels across surfaces.

So the competitive line has moved from price to authenticity, and own-label has won the authenticity argument at the category level while remaining entirely absent from the need-state level. Both halves matter. Branded manufacturers cannot dismiss own-label as the cheap option. Retailers cannot assume the credential they have built transfers to the question the anxious owner asks.

Where branded still wins

Branded holds where the credential is hardest to replicate: therapeutic and veterinary diets, specialist functional formats, and the premium end where formulation IP and clinical substantiation are genuine barriers. Mintel notes leading branded dog food values under pressure from premium challengers, but the vet-diet space remains branded-dominated for structural reasons — and Section 3 shows those same structural reasons now govern who the machine names.

8. The UK problem nobody has named

Everything above has been about behavior. This section is about law, and it must be exact, because precision here is the difference between an original insight and a correction in the next issue.

What actually governs this

In Great Britain, pet food marketing and labelling is governed by assimilated Regulation (EC) No 767/2009 — retained EU law, still in force. The Chartered Trading Standards Institute's government-backed Business Companion guidance, reviewed May 2026, is explicit that this assimilated law continues to apply, and the Food Standards Agency references it directly.

The enforcement structure is not where most people assume. The FSA is the central competent authority for feed policy and safety. Local authority Trading Standards enforces labelling and marketing day to day. APHA handles animal by-product approvals. DEFRA sponsors policy but does not enforce the label. The FEDIAF Code of Good Labelling Practice is influential and widely referenced, but it is self-regulatory interpretation, not binding law.

The declaration that decides whether it is a meal

The “type of feed” declaration — complete or complementary — is required under Article 15(a), and Article 14(3) carries it onto distance-selling material: the mandatory particulars must appear there, or be supplied before the contract concludes. Exactly four may be deferred to delivery — the operator's name and address, net quantity, minimum storage life, and batch number. Type of feed is not among them.

This is worth stating plainly because the opposite is widely assumed. The declaration is not a pack-only obligation that listings may omit at their discretion. It is required online.

So the rule exists. What does not consistently exist is the execution. In practice the declaration is frequently absent from listing text, buried in a pack shot, rendered inside an image, or parked in a compliance block a machine will not associate with the product's identity.

And the probe shows the consequence with uncomfortable clarity. Across the entire captured corpus of AI answers to UK pet-nutrition questions, the word “complementary” appeared once. “Complete and balanced” surfaced as a named attribute in eight of 106 answered runs. The single most consequential nutritional fact in the category — whether this food is the animal's diet or an addition to it — has effectively not made it into the machine's working vocabulary.

The most important nutritional fact about a pet food is legally required to be on the listing, and a machine reading a UK listing still usually cannot tell whether the food is a meal.

The machine is reaching for the wrong rulebook

Here is a finding we did not go looking for, and it is the sharpest regulatory observation in the study.

The machines answer UK questions using the American regulatory frame. Asked which UK dog foods use human-grade ingredients, one assistant explained the standards used by the Association of American Feed Control Officials — a US body with no standing over a UK label. AAFCO appears in the captured UK answers explaining UK products, alongside FEDIAF, whose code actually governs UK labelling. On the retailer surface specifically, AAFCO was the only one of the two to appear in the captured answer text at all.

We are not claiming the American frame dominates — across the full capture the two bodies appear at broadly similar frequency, and our stored excerpts are too short to establish a ratio. The point stands without one: the regulatory vocabulary being repeated back to UK owners is frequently not the one their products are labelled under.

Composition: legal, and machine-useless

Reg 767/2009 permits ingredients to be declared by category — “meat and animal derivatives,” “cereals,” “oils and fats.” Lawful, long-established, defensible in formulation terms. Also close to unusable for a machine trying to answer “does this contain chicken.” The FEDIAF Code sets the declared-percentage thresholds behind “with,” “rich in,” and “flavoured with,” which adds precision for those who know the convention and none for a reader who does not.

Category naming is not a compliance failure. It is a machine-readability failure that compliance permits.

The claims corridor, precisely

Products presented as treating or preventing disease require authorization under the Veterinary Medicines Regulations 2013 — whether medicinal by presentation or by function. The VMD's guidance on unauthorised products, “Words and Phrases,” version 11, reviewed 25 November 2024, sets out the line. Specific reference to a named condition — arthritis, cystitis, anaemia, Cushing's — is prohibited. Incorporating an active ingredient from an authorized veterinary medicine at an efficacious level makes a product medicinal by function regardless of description.

The permitted formulation, per VMD and BSAVA guidance, is that a product may support or maintain normal health in a healthy animal. It may not improve or boost health, or prevent disease. In January 2026 the VMD published clarifying guidance on advertising non-medicinal veterinary products, restating that the law has not changed and claims must be limited to health maintenance in healthy animals.

Now recall Sections 3 and 4. The highest-demand need-states sit directly against this boundary — and they are the exact queries where the machine collapses to a clinical shortlist. The brands winning that shortlist are, disproportionately, the ones whose therapeutic ranges are authorized to say the things everyone else cannot. That is not a content deficiency. It is a structural feature of the UK market, and any content strategy that ignores it is briefing writers to produce non-compliant copy in pursuit of a position the law reserves for someone else.

“Natural,” “human grade,” and the definitional hole

“Natural” is a FEDIAF term — self-regulatory, not statutory.

“Human grade” has no UK legal definition for pet food at all. The ambiguity runs deep: virtually any meat-containing pet food could claim human-grade ingredients, because most ingredients begin human-grade, while common processing such as rendering removes that status — meaning no finished pet food is human grade at point of sale in the sense a consumer would understand.

We are not going to pretend that is resolvable. The ambiguity is the finding — a term with no legal definition, appearing in premium positioning across branded and own-label alike, being read by a machine that treats it as a specification and explains it using American rules.

Advertising and environmental claims

The CAP Code applies to marketing communications including retailer product pages; comparative and free-from claims require substantiation. “Vet recommended” sits at the strictest end.

The environmental-claims regime changed materially and recently. The consumer-protection provisions of the Digital Markets, Competition and Consumers Act 2024 came into force on 6 April 2025, giving the CMA power to investigate and fine breaches of consumer law directly, without going to court — up to 10% of global annual turnover. The Green Claims Code sets the substantiation standard, and environmental claims are a stated CMA enforcement priority.

We looked for dated ASA or CMA action specifically on pet food environmental claims and found none. That is an evidence gap, not a clean bill of health.

9. What to do next

Nine sections of argument earn a short list of instructions. These are staged in order of confidence, and each is labeled with how strong the evidence beneath it is — because a recommendation pitched above its evidence is how trade readers learn to stop believing you.

1. Put the type-of-feed declaration in machine-readable text on every listing. Evidence: strong — law plus first-party. Legally required online under assimilated Reg 767/2009 Art 14(3), and the probe found “complementary” appearing once in the entire captured corpus. This is the cheapest fix on the list and the largest gap. Start here.

2. Work out whether you are in the shortlist, and at which questions. Evidence: strong — first-party. Section 3's collapse finding is the single most actionable result in the study. Take your ten highest-value need-state questions, run them, and record whether you appear. Do it three times, because a third of our repeated runs returned a different lead brand. Broad brand-trust visibility tells you nothing about this.

3. Replace inference with statement on life stage, breed size, and format. Evidence: strong. Anything currently carried only by a photograph is invisible to two thirds of your audience.

4. Build the credential, and build it where the machine reads. Evidence: strong on mechanism, inferential on remedy. Brand-owned content was 5.9% of citations; reviews and forums were 54%. Owned content is not where this is won. What the machines repeat is third-party ratified credentials. That points at clinical substantiation, publishable evidence, and presence in the review and comparison estate rather than at more brand copy — though we should be honest that we have measured the problem far more precisely than the fix.

5. Audit every functional claim against the VMD line. Evidence: strong. Rewrite toward “supports” and “maintains” rather than deleting; the compliant sentence usually exists.

6. Contest the life-stage and transition territory now. Evidence: moderate — first-party, small n. Life-stage questions produced almost no brand mentions in our study. That is the clearest open ground the probe found, and it is bounded by a small sample.

7. Treat Q&A and reviews as the corpus the machine actually reads. Evidence: strong on citation share, moderate on conversion effect. The underlying content-to-conversion evidence is US and vendor-commissioned; the citation share is ours and unambiguous.

8. Validate any content brief against each retailer's actual specification. Evidence: strong on necessity, gap on the specs. Grocery and specialist retailer content specs are not publicly published. Get them from the retailer.

9. Audit environmental and “human grade” claims for exposure. Evidence: strong on the law, thin on precedent. DMCC powers live since April 2025 with serious penalty exposure; no pet-food-specific enforcement located, which is not the same as safety.

10. Keep the listing accurate through reformulation and pack change. Evidence: mechanism sound, frequency unmeasured.

Where the evidence runs out

Five things are weaker than the rest, and naming them is what buys the right to be believed on everything above.

We were wrong about the premise, and the probe is why we know. We predicted brands would vanish at the need-state layer. They do not. Had we published on the original thesis we would have been confidently, checkably wrong. The finding that replaced it is better — but it is worth saying plainly that the study's most valuable output was refuting the thing we set out to prove.

One surface did not report. Gemini returned no usable answer in 53 of 54 runs, captured from a non-UK IP with UK scope forced by prompt. That is an instrument limitation, not a finding about Gemini, and none of the analysis above rests on it. Google AI Overview rows are API-captured and “AI box appeared” should be read as directional.

Several blocks are small. Life-stage, provenance and subscription each produced only two or three answered runs. Those findings are flagged as directional and should not be quoted as rates.

The pet-specific AI adoption figure does not exist. We measured what the machines say. Nobody has measured how many UK owners are listening.

Some of the evidence is ours. The retail-AI recommendation study and the page-level claim testing in Sections 3 and 5 are Genrise research, cited by the company that produced them in an article that recommends acting on them. We have labeled them first-party wherever they appear, exactly as we labeled the brand-commissioned veterinary research we used in Section 2. Neither has been independently replicated. Readers should weigh both accordingly.

Personalization is untested. Every Rufus capture came from a fresh account with no purchase history. The agentic shelf we measured is the open one. What happens once account memory has something to remember is the obvious next study, and we have not done it.

And what would falsify what is left

If a repeat of this study six months from now shows the need-state answer set widening rather than holding — more brands, lower concentration, own-label crossing into functional territory — then the collapse is a transient artifact of early models and this article overstates it. That is a testable claim on a known instrument, and we would rather it were tested than believed.

Method note

Fieldwork August 2026. Thirty prompts written in natural UK shopper English across seven blocks: vet displacement, private-label parity, functional need-states, life-stage, provenance and trust, subscription, and a control set. Twelve Tier 1 prompts run at three repetitions across four conversational surfaces plus Amazon.co.uk Rufus; eighteen Tier 2 prompts at lower repetition. 180 conversational captures, 54 Rufus captures, 1,785 citations logged with source-owner classification. All answers dated and screenshotted.

Reporting posture: category-level patterns only. No individual brand's visibility is reported as a score, and no brand is singled out for weakness. Where brand groups are described, the description covers brands from all three major manufacturers in the category, including the co-author's. Genrise scoring methodology is not disclosed.

Limitations are stated in full in Section 9.

Sources

First-party

• Genrise UK AI Probe, fieldwork August 2026 — 30 prompts, 234 captures, 1,785 citations (UK-specific, first-party)
• Genrise retail-AI recommendation study — 610,096 recommendation events, 22,824 products, 8 categories, 3 markets (cross-category, not pet-specific; first-party)
• Genrise page-level claim-retrieval testing, wet dog food — 878 US and 5,762 UK question-by-page judgments, 122,000+ phrase checks (single category, single retailer surface; first-party)

Regulatory and government

• Assimilated Regulation (EC) No 767/2009, Articles 14(3) and 15 — legislation.gov.uk
• Veterinary Medicines Regulations 2013; VMD, Unauthorised Products — Words and Phrases, v11, reviewed 25 November 2024; VMD guidance on advertising non-medicinal veterinary products, January 2026
• Chartered Trading Standards Institute, Business Companion, “Retail sale of pet food,” reviewed May 2026
• Food Standards Agency, animal feed guidance, March 2025
• Competition and Markets Authority, Veterinary Services Market Investigation final report, 24 March 2026; draft Order, consultation opened 21 July 2026
• Digital Markets, Competition and Consumers Act 2024, consumer provisions in force 6 April 2025; CMA Green Claims Code
• CAP Code; FEDIAF Code of Good Labelling Practice (self-regulatory)

Market and category

• UK Pet Food, pet population research, Kantar fieldwork January 2026, n≈8,951 (UK-specific)
• Mintel, UK Pet Food and UK Pet Retailing market reports, 2024–2025 (UK-specific, commercial)
• Circana, European private-label pet food data, 2022–2023 (UK-inclusive, commercial)
• Ken Research, UK Pet Food Market (UK-specific, commercial)
• Pets at Home Group plc, FY2026 results, year ended 26 March 2026 (company disclosure)

Consumer and behavioral

• Dogs Trust, National Dog Survey 2024, n>406,000 (UK-specific)
• YouGov, family pets attitudes (UK-specific, panel)
• PDSA, PAW Report 2025 and prior waves (UK-specific)
• Ofcom, Online Nation 2025, 10 December 2025; Adults' Media Use and Attitudes, 2025 (UK-specific, regulator)

Veterinary and nutrition authority

• PLOS One, UK and Irish veterinary student nutrition information sources, fieldwork October 2023–January 2024 (UK-specific, academic)
• Canadian Veterinary Journal, pet food purchasing determinants, n=2,181 (Canada, proxy)
• Purina Institute / Forward Group, pet nutrition trust survey, June 2023 (US, proxy — brand-commissioned)

Proxy-buyer mechanism

• Froley et al., Maternal & Child Nutrition, 2025, infant formula online marketing (US, proxy)
• Foods, infant milk formula purchase intention (China, proxy)

Platform disclosures

• Amazon Q3 2025 earnings call, November 2025; Q4 2025 earnings, February 2026 (global, company disclosure)

Declared gaps — no source located

• UK pet-specific AI-use rate for nutrition advice
• Subscription as a share of UK online pet spend
• UK online pet sales growth relative to category
• Independent (non-Genrise) UK pet-specific content-to-conversion evidence
• Retailer content specifications for UK grocers and pet specialists
• UK pet food reformulation, resizing and delisting frequency
• UK pet food switching rates and inflection-point triggers
• Dated ASA or CMA enforcement on pet food environmental claims