Why ChatGPT won’t recommend your vegan brand—and the proof it needs

“Why doesn’t ChatGPT recommend my vegan brand?” I hear a version of that question from almost every vegan founder I work with. The answer is structural, not personal: the model can’t verify your claims, so it recommends brands someone else already has.

ChatGPT, Perplexity and Gemini don’t recommend a vegan brand because its website says it’s vegan. They recommend brands whose claims a third party can verify: a certification body, a lab report, a press outlet, a retailer’s structured product data. A self-asserted “vegan”, “cruelty-free” or “sustainable” is unverifiable marketing copy the model must hedge or quietly exclude. A certified claim is a citable fact—and citable facts are what AI answers are made of.

The question founders actually ask

Here’s the exercise I give every founder who asks me this. Open ChatGPT, Perplexity and Gemini, and ask each one for the best vegan snacks in Berlin, Vienna or Chicago—whichever city you actually sell in. Three models, one question, no ad slots. There is no auction to win your way into; nothing you bid on gets you there. What comes back is each model’s best defensible picture of your category, assembled from sources it can check.

In audits like this, the pattern we usually see is consistent: the answers lean on listicles, retailer category pages and press coverage; the brands that appear are the ones whose claims live outside their own site; and the wording is hedged—“popular options include…”, “some well-known brands…”. Results shift week to week and model to model, so a single snapshot proves little. The useful signal is whether you appear at all, and what the model can point to when it names you.

Why LLMs hedge on ethical claims

OpenAI’s Model Spec—the document that tells ChatGPT assistants how to behave—ranks possible answers like this:

confident right answer > hedged right answer > no answer > hedged wrong answer > confident wrong answer

Read that ranking again, because it explains everything. A model would rather say nothing than assert something it can’t back. Now look at your homepage: “100% vegan”, “cruelty-free”, “sustainably sourced”. On your site, those are values. To a language model they’re assertions with no external anchor—the same epistemic status as “the world’s best coffee”. And the spec tells the assistant exactly what to do with them.

• Hedging. The model includes you but softens it: “I think”, “It might be”, “many people consider”. Your brand appears—framed as an opinion.
• Disclaiming. It attributes the claim instead of asserting it: “There are reports suggesting…”. Your positioning becomes somebody else’s hearsay.
• Silent exclusion. The safest option: it recommends the competitor whose claims a retailer, a certifier or a journalist already documented, and never mentions you at all.

The fix in one sentence

Here it is: make every load-bearing claim verifiable by someone who isn’t you, then make that verification machine-readable. That’s the whole game. The rest of this article is the how, the evidence, and the checklist.

The strongest vegan signals are third-party certifications a model can look up. The Vegan Society’s Vegan Trademark bills itself as the leading third-party vegan certification scheme in the world, with more than 70,000 certified products—and a dedicated team checks each application. In Europe, the V-Label—an internationally registered seal run under the European Vegetarian Union—appears on more than 70,000 products from over 4,800 licensees. Both give your claim an issuer, a registry and a lookup path.

• Lab tests: allergen and contaminant results from an accredited lab, especially for free-from and “may contain” claims
• Third-party registries: certifier product databases and retailer ingredient pages that mirror your claim
• Press verification: a journalist or outlet stating the claim, which models treat as independent attestation

Make the proof legible to crawlers

A certification logo is for humans. For the crawlers that feed AI answers, schema.org defines a Certification type—“an official and authoritative statement”—attached to a Product (or Organization, Service, Place) via the hasCertification property, with issuedBy for the certifier and certificationStatus to mark it active or inactive. This is a documented standard, not an SEO trick. Minimal version for a product page:

{
  "@context": "https://schema.org",
  "@type": "Product",
  "name": "Sea Salt Snack Crackers",
  "hasCertification": {
    "@type": "Certification",
    "name": "Vegan Trademark",
    "certificationStatus": "active",
    "issuedBy": {
      "@type": "Organization",
      "name": "The Vegan Society"
    }
  }
}

In every values-brand audit I run, the first two things I check are whether the certificate exists as data—somewhere a crawler can read it—and whether the claim wording is identical everywhere it appears. Those two checks surface most of the gap.

• Product pages: hasCertification markup, not just a logo image next to the word “vegan” in a paragraph
• About page: organization-level claims, marked up and worded exactly as on product pages
• Retailer pages and your Google Business Profile: the same claim, in the same words—not a cousin of it

The evidence this week: claim substance moves appearance

This stopped being theory recently. On September 22, 2026, SPINS—the retail analytics firm—launched Agentic Discovery, a platform for optimizing consumer brands’ AI visibility. SPINS’ chief product officer framed the shift plainly:

The next battleground for consumer brands isn’t the shelf or the search engine, it’s the AI recommendation.

Two case studies from the platform show what actually moves appearance rates. MadeGood, a snack brand, closed content gaps on purchase-driver claims—school-safe, nut-free—using structured product data, and gained 12.5 points of appearance rate in ChatGPT, hitting 81% appearance for school-safe queries and a 49% increase for nut-free queries. Death Wish Coffee deployed structured content around its fair-trade claims and gained 13.9 points in Google AI Overviews, reaching 100% appearance on high-caffeine queries and a 50% increase on fair-trade queries.

Notice what neither brand did: chase keywords. They documented claims buyers actually ask about—claims that needed substantiation—and appearance followed.

The own-label wall—and the small-brand rebuttal

The counterargument is that only big brands win in AI answers, and there’s data behind it. EMARKETER’s AI Visibility Index for food and beverage analyzed 7,460 ChatGPT recommendations across 16 US categories (July 2026). The No. 1 brand overall was Great Value—Walmart’s own label—with a mention rate of 5%. Retailer brands dominate, and that’s the wall.

But read the second number. Five percent. The single most-recommended brand in the entire index appears in one in twenty recommendations—visibility is fragmented enough that nobody owns the answer. And the GPT-5.5 to GPT-5.6 transition tightened it further: unique brands mentioned fell just 2%, while unique domains cited dropped 17%, with Walmart by far the most-cited source for savory-snack recommendations. The model narrowed its source pool. Being one of the defensible, well-documented sources matters more now, not less.

Here’s the rebuttal for a five-person vegan brand: you are not trying to win “best snacks”. You’re trying to own “nut-free vegan snacks for school lunches”—the exact kind of niche query where MadeGood sits at 81% appearance. Specialized claims, documented properly, are the terrain where small brands beat own labels, because own labels don’t bother documenting them.

EmpCo makes it double jeopardy

If you sell into the EU, the substantiation question is no longer optional. The EmpCo Directive’s rules apply from September 27, 2026, and the European Commission’s FAQ addresses vegan labels directly. Asked whether a “vegan” label counts as a sustainability label, the Commission’s answer is case by case.

In practice: “vegan” as a plain ingredient claim generally stays outside the sustainability rules; pair it with planet or animal-welfare framing and it becomes a regulated environmental claim that needs certification and substantiation. Which is exactly what an LLM asks for. Regulators and models have converged on the same standard—so trust-signal work is compliance work, and compliance work is trust-signal work.

When incidents outrank values

In September 2026, Miyoko’s Creamery—the recalling firm is Prosperity Organic Foods—initiated a Class I recall of a single lot of its European Style Cultured Vegan Salted Butter over potential E. coli contamination. Grocers pulled the lot; at the time of reporting, the FDA hadn’t even posted the recall. A separate August 2026 case: Frankie’s Organic issued an allergy alert on its Plant-Based Vegan Cheddar Puffs—best by “05 16 2027”—after two complaints of allergic reactions, because the “vegan” product may have contained undeclared milk.

Two different failures, one shared mechanic. Safety incidents get exhaustively documented—retailer notices, news coverage, lot numbers—and that documentation is exactly what crawlers and models love to cite. A recall can enter AI answers and overwrite years of positioning, because the incident is better-documented than the claims were.

That’s the defense case for everything above. If your certifications, lab tests and claim data are structured, consistent and crawlable, an incident becomes one fact among many in your record. If your claims exist only as adjectives on a homepage, the incident is the single best-documented fact about you—and models will use it.

The six-step trust-signal checklist

1. Certify. Put your load-bearing vegan claim in the hands of a body a model can look up: the Vegan Trademark, the V-Label, or an equivalent certifier in your market. One scheme, done properly, beats three logos nobody can verify.
2. Structure. Mark the certificate up with hasCertification on every product page, and carry organization-level claims on your about page. A logo image is invisible to a crawler.
3. Substantiate. Keep an evidence file per claim: the certificate PDF, the lab report, the press link. If a claim has no artifact behind it, decide whether to certify it or stop making it.
4. Syndicate consistently. The exact same claim wording on your site, your retailer pages and your Google Business Profile. Contradictory wording between surfaces reads as hedging to a model.
5. Monitor. Track how ChatGPT, Perplexity, Gemini and Google AI Overviews describe your brand week over week—a tracker like Cited does this in real time and turns the gaps into to-dos. You can’t fix what you never see.
6. Re-verify. Certificates expire and formulas change. A certification marked “active” that has lapsed is worse than none at all.

How we audit this

Everything above is checkable, which is why we audit it at a fixed price rather than a retainer. A values-brand audit covers: the queries your buyers actually ask AI in your category; where you and your competitors appear, model by model; a claim-by-claim inventory of every load-bearing claim and what backs it; markup and consistency gaps across your site, retailer pages and Google Business Profile; and a prioritized fix list—what’s a five-minute change and what’s a certification project.

The AI Visibility Audit is a €99 flat rate, because trust-signal work should be checkable, not open-ended. If you want to see how this applies to your brand specifically, that’s the place to start—or write to us and tell us what your claims are. We build for vegan brands, and this is the core of that work.

None of this is keyword SEO, and that’s the point. The models have already decided they won’t take your word for it. Stop asking them to. Give every claim an issuer, a date and a lookup path—and watch the hedging stop on its own.