AI Visibility: Why 4 Engines Score One Brand 10 to 100

Same brand, same skincare query, same afternoon: ChatGPT scored Tatcha a 10 out of 100 and Claude scored it a perfect 100. Not a rounding gap - the opposite answer, from two engines your customers treat as interchangeable.

I did not expect that. I ran our own AI-visibility audit - 31 direct-to-consumer brands, three categories, four engines (ChatGPT, Claude, Gemini, Perplexity) - mostly to stress-test our own scoring, expecting the engines to broadly agree on who the strong brands are. They did not. Not even close.

From what I have seen across our audit data, the gap between engines is not noise around some true score. Check a single engine and you have measured a quarter of your AI visibility, not all of it. Pick a different one and the same brand goes from invisible to dominant.

Every AI engine gives the same brand a different visibility score

Here is the mean visibility score per engine, across the 20 brands we ran against all four. Same brands, same category queries, same day. The only thing that changed was the engine:

Claude hands out visibility at nearly three times the rate ChatGPT does. That is not a rounding gap. ChatGPT was the stingiest engine for 16 of 20 brands; Claude the most generous for 14 of 20. They are not measuring the same thing - ChatGPT guards a short list, Claude recommends generously.

Now the number that actually matters. Take each brand's four engine scores and ask a simple question: do they land in the same rating tier - Low, Limited, Moderate, Strong, Excellent? Every single one of the 20 brands lands in a different tier depending on which engine you ask. Twenty out of twenty. There is not one brand in the set where the four engines agree on the tier.

“

Large companies keep announcing newer and larger gen AI models, but how do you know which model is truly the best if your evaluation methods aren't accurate or well studied?

One brand's AI visibility score swings from 10 to 100

Zoom into individual brands and it gets sharper. The average gap between a brand's best and worst engine in our set was 36 points. Here are the brands the engines fought hardest over:

Tatcha is the one that stops you. On ChatGPT it scores 10 - effectively invisible. On Claude it scores 100 - the only thing the model wants to talk about. Same brand, same skincare query, same afternoon. If Tatcha's team checked ChatGPT they would panic; if they checked Claude they would declare victory. Both would be wrong, because both measured a quarter of the picture.

This is the brand-recognition lottery we documented in the audit preprint, now showing up in our own first-party numbers. One engine is not a measurement. It is a draw - which is why one audit isn't enough.

Schema and AI crawler access did not explain AI visibility

Here is the part I did not expect, and the one that should change where you spend money. I checked whether the on-site signals everyone obsesses over - homepage structured data, AI-crawler access - explained any of the score. They did not.

• Brands with homepage schema averaged 36. Brands with none averaged 35. A one-point difference, which is to say no difference.
• Zero of the 31 brands blocked GPTBot, ClaudeBot, PerplexityBot or Google-Extended. Crawler access was not the lever either - everyone already allows them.

So if it is not your schema and not your robots.txt, what is it? It is how densely your brand is referenced across the web these models trained and retrieved on. That is the whole thesis of our Mention Density Model: AI recognition is earned off your site, in the corpus, not on your homepage. Our own audit just failed to find any on-site signal strong enough to argue otherwise.

This is not permission to ignore schema. Clean structured data still fixes how engines describe you once they know you exist. But it will not get you into the answer. Being in the answer is a mention-density problem, and you solve it off your own domain.

The wider research lands in the same place. The Princeton GEO study (Aggarwal et al., 2023) tested what actually moves visibility inside generative engines and found the winners are content-credibility signals - citing sources, adding statistics, working in quotations - lifting visibility up to ~40% for the strongest. Notice what is not on that list: homepage schema.

Category and brand size shape AI visibility, not effort

Category matters more than effort. Mean visibility ran Sports & Fitness 47, Cosmetics & Beauty 35, Food & Nutrition 25. Functional food and supplements is simply a harder room to get recommended in - the engines hedge more on anything health-adjacent, and it shows up as a flat 20-point category penalty.

And real-world fame is not AI visibility. Lululemon and Nike saturate at 100. But Adidas sits at 43, Under Armour at 33, and serious DTC names - Vuori, Rhone, Fabletics - sit at 17, near-invisible despite real revenue and real customers. AI keeps a short, sticky list, and being big offline does not buy you a seat at it.

How to measure AI visibility across four AI engines

One audit, one engine, one snapshot is how almost everyone measures AI visibility right now. Our own data says that is close to worthless. Here is what I run instead, and you can do it for pennies:

1. Score all four engines, never just one. ChatGPT, Claude, Gemini, Perplexity. A combined score is fine - the trap is trusting a single-engine reading, or a blend that hides the tier-level disagreement we found on every brand tested.
2. Read your worst engine, not your best. A 100 on Claude and a 10 on ChatGPT is a ChatGPT problem, not a trophy. The floor is the work.
3. Spend off-site, not on-site. Schema and crawler access explained none of the gap in our data. Mentions, reviews, third-party coverage and entity density did. Put the budget where the signal is.
4. Pick category-aided queries you can actually win. Your category sets the ceiling. Find the rooms where mid-market brands get recommended and own those, instead of chasing "best [product]" terms the incumbents have locked for a decade.
5. Re-run it. One snapshot is a mood, not a measurement. Engines drift between runs, so track the trend across several reads, never a single snapshot. The research backs this up: Schulte et al. (2026), "Don't Measure Once" argues AI-search visibility has to be measured repeatedly, and multi-prompt LLM evaluation research shows how sharply model results shift across prompt variations.

None of this is theory about how the models work. It is just what our own 31-brand audit made impossible to ignore: the engines do not agree, the disagreement is the signal, and the lever is off your website. Measure it that way and your visibility score finally means something, instead of swinging on whichever engine you happened to open.

Want a visibility score built on all four engines, with the per-engine breakdown behind it? Run GEOlikeaPro's Visibility checker - it scores ChatGPT, Claude, Gemini and Perplexity separately and rolls them into one number you can trust, because you can always see the spread underneath it.

Latest brands checked with our AI tools

The newest public reports from our AI visibility audit, hosting checker, and AI readiness checker - run yours to add your brand.