State of AI Visibility 2026: AI Already Knows Your Brand. That Isn't the Problem.

Every week, Surfaceable asks three AI platforms the same question about a few hundred established brands: "What is [brand]?" We expected the results to show which companies AI models know well and which they don't.

They showed something more useful. For established brands, recognition is universal — and every brand that looked invisible to AI turned out to be a flaw in how it was being measured.

The live numbers are in the State of AI Visibility report, which recalculates as new checks run. This post explains what's behind them, as of late October 2026.

Key findings

• AI models recognise every established brand we track. After fixing our measurement and re-checking all 303 brands on 26 October, ChatGPT (GPT-4o mini), Claude (Haiku 4.5) and Perplexity (Sonar) each recognised all 303.
• Before cleaning, the same data told a very different story. Raw results had Claude recognising just 57.8% of brands and 4.5% of brands unknown to every platform. Both figures were artefacts.
• The artefacts came from three places: a retired model ID whose errors were scored as "not mentioned", brand names stored in a form nobody writes, and punctuation.
• Structured data is the most common gap. Half of the audited homepages have no JSON-LD structured data at all, and 58% don't describe their own organisation in schema — even though the average SEO score is a healthy 87.4/100.

Finding 1: Recognition isn't the gap for established brands

The prompt we use is deliberately simple: "What is [brand]? Give me a brief overview: what they do, who they serve, and whether they are well-known or recommended in their space." A brand counts as recognised when the answer names it and describes it substantively, rather than saying the model has no information.

After fixing the problems described below, we re-checked all 303 brands across 50 industries on the same day:

Platform	Model	Brands recognised
ChatGPT	GPT-4o mini	303 of 303 (100%)
Claude	Haiku 4.5	303 of 303 (100%)
Perplexity	Sonar	303 of 303 (100%)

Not one brand went unrecognised. In earlier weekly checks a handful had: Lloyd's of London was once missed by Perplexity, and McDonald's UK, St. James's Place and IVC Evidensia were missed because of how we'd stored their names. On a clean re-check, every one was recognised by all three platforms.

The practical takeaway: if you're an established company, "does ChatGPT know who we are?" is the wrong question. It almost certainly does. The questions that matter are whether AI recommends you when someone asks for the best option in your category, how it describes you compared with competitors, and which sources it relies on to do that. Recognition is the floor, not the goal.

That doesn't hold for newer or smaller brands. Every brand in this benchmark is an established company with years of press coverage, reviews and web presence. A two-year-old startup has far less of that for models to learn from, and its recognition rate will be lower.

Finding 2: Most "AI visibility gaps" are measurement errors

Here is what the raw data said before we cleaned it, next to what's true:

Metric	Raw data	After removing invalid checks
Claude recognition rate	57.8%	100%
Brands no platform recognised	4.5%	0%
Brands flagged "losing AI visibility"	121	Trend reset — the drop was an outage

Each gap had a cause that had nothing to do with the AI models' knowledge.

1. A retired model, scored as silence

From the last week of September, Claude's recognition rate in our data fell from between 87% and 100% to exactly 0%. Claude hadn't forgotten every brand overnight. The model version we were calling had been retired by the API provider, every request failed, and our code recorded each failure as "not mentioned".

For four weekly runs, 120 brands lost a third of their apparent AI presence, and 118 of them were flagged as "losing" visibility. On our own public brand pages. We've since switched to Claude Haiku 4.5, removed the invalid Claude results from those snapshots, and changed the code so a failed request is excluded rather than counted as a miss.

When we re-ran the prompt for brands Claude had supposedly missed — GitLab, Greggs, Gymshark, Freshworks, Grant Thornton, Hammerson — it gave a detailed, accurate overview of every one.

2. Names nobody writes

Some brands in our dataset had been named automatically from their domain: "Britishgas", "Cliffordchance", "Taylorwimpey". A model asked about "Britishgas" usually answers about British Gas, with a space — which a simple text match doesn't count as a mention. A few brands had also been added twice under different names, such as "Clifford Chance" and "Cliffordchance".

We've given these brands their proper names, merged the duplicates, and changed the matching to ignore spacing and punctuation. Re-checked under their real names, every one of them was recognised by all three platforms.

3. Punctuation

"McDonald's UK" with a straight apostrophe doesn't match "McDonald's" with a curly one, which is what models often write. "St James Place" doesn't match "St. James's Place". Small differences like these were behind most of the remaining misses.

Why this matters if you track AI visibility

Every AI visibility tool — ours included — reduces a paragraph of generated text to a yes or no. Before you act on a low score from any of them, ask:

• What happens when a request fails? If errors count as "not mentioned", a model outage looks like a collapse in visibility.
• How are names matched? Check that your brand is tracked under the name people and models actually use, including common variants and abbreviations.
• Is a sudden drop across many brands at once? Real visibility changes happen brand by brand. A drop across everything on the same day is almost always a pipeline problem.
• Which model is being asked? Results from a small, cheap model (which most tools use to keep costs down) can differ from the flagship model your customers use in the app.

Finding 3: Technical SEO is strong but uneven

We also audit each brand's homepage weekly — on-page SEO, structured data and accessibility. Across the 150 brands with a current score:

• The average score was 87.4/100, with a median of 90.
• 55% scored 90 or above, 38% scored 75–89 and 7% scored 50–74. None scored below 50.
• The highest-scoring sectors were marketing software, sports and fitness, B2B software and charities. The lowest were ecommerce brands, healthcare providers and travel. Industry samples are small, though — about three audited brands each — so treat sector rankings as indicative rather than conclusive.

The headline scores hide a consistent gap: structured data. Of the 150 audited homepages:

Issue	Share of brands
No WebSite or WebPage schema	68%
No Organization schema	58%
No JSON-LD structured data at all	50%
No skip-navigation link	43%
Missing canonical URL	23%
Missing meta description	22%

Structured data is how a site states facts about itself in a form machines don't have to guess at — who the organisation is, what the page is, where its official profiles live. Half of these brands give search engines and AI systems none of it, and rely entirely on what others say about them.

SEO score said nothing about AI recognition here: every brand was recognised, whatever its score. That's a ceiling effect, not proof that SEO doesn't matter for AI. When everyone is recognised, recognition can't distinguish good SEO from bad. The places where technical SEO is likely to matter more — whether AI search engines can crawl, cite and recommend your pages — are what we'll measure next.

What to do with this

1. Stop worrying about whether AI knows you exist if you're an established brand. Test whether it recommends you instead: ask the questions your customers ask ("best [category] for [use case]") and see who gets named. Our free AI Prompt Tester runs the same prompt on ChatGPT, Claude, Gemini and Perplexity.
2. Audit your measurement before your marketing. If a tracking tool shows a sudden drop, check for errors and name mismatches before rewriting content.
3. Add structured data, then fix the other basics. At minimum, Organization and WebSite schema on your homepage — our free schema generator builds the Organization markup. Then check titles, descriptions and canonicals with the free meta tag analyser.
4. If you're a newer brand, recognition is still your first job. Earn the third-party coverage — reviews, comparisons, press, community discussion — that models learn from.

Methodology

• Brands: 303 established companies across 50 industries, with a UK focus. Each has a brand page with its latest results.
• AI checks: the prompt above, sent to GPT-4o mini, Claude Haiku 4.5 and Perplexity Sonar. The recognition figures come from a full re-check of all 303 brands on 26 October 2026, after the fixes described above. The raw figures in Finding 2 come from weekly checks between 12 August and 21 October 2026.
• Data corrections: around 40 brands were renamed to their real names, two incorrect domains were fixed, and five duplicate brand entries were merged.
• SEO scores: the average of Surfaceable's on-page SEO, structured data and accessibility checks on each brand's homepage.

The live report updates as new checks run, so its figures will move on from the ones in this post. You're welcome to cite either, with a link.