Howseen: 85.5% of Brands Were Never Named on Their First AI Visibility Scan

Most AEO dashboards show movement after months of work. Howseen’s October 2026 pooled study asks a harder baseline question: where do brands start before anyone optimizes?

The answer is blunt. On each brand’s first completed scan, 307 of 359 brands (85.5%) were never named in any answer to non-branded buyer questions. Only 2.9% of all 6,923 answers named the brand at all. ChatGPT named 11.3% of brands at least once; Gemini 8.5%. Own-domain citation was rarer still — 10.6% of brands — but when it happened, naming usually followed.

That is not a forecast that GEO is impossible. It is a reminder that category recommendation is a high bar, and that first-scan baselines should shape how you set expectations and sequence work.

What Howseen measured

Howseen took the first scan for each brand between August 26 and October 20, 2026 — before follow-up optimization cycles, by design.

Named means the answer text mentions the brand. Own site cited means a source URL contains the brand’s domain.

Google AI Overviews, AI Mode, and Perplexity appear only on paid plans in Howseen’s product, so sample sizes for those engines are small (22–50 brands). The published headline comparisons focus on ChatGPT vs Gemini.

Caveats the authors stress: mostly French SMBs (325 of 368 language settings), median ~5 questions per brand, one scan per brand, non-deterministic answers, and correlation ≠ causation. The study describes a starting line, not a controlled experiment.

Finding 1: Invisibility is the default first outcome

Among brands with enough answers to stabilize counts slightly, four in five still never appear. That aligns directionally with enterprise studies showing concentrated AI visibility — see Sentic’s October 2026 report — but Howseen’s sample is the long tail: local services, trades, and mid-market sites that are not national household names.

Implication: “We are not in ChatGPT yet” is normal on day one. The strategic question is whether you are retrievable and citable on the questions buyers ask before you expect a name drop.

Finding 2: ChatGPT names slightly more brands than Gemini

The gap is modest but consistent with other October datasets showing engine-specific mention rates — not one blended “AI rank.” Google’s own documentation warns that AI Mode and AI Overviews may use different models and show different links; treat each surface separately, as in Why Each AI Engine Reads a Different Search Index.

Finding 3: Own-site citation and naming move together

Only 38 brands had an own-domain citation on the first scan. 34 of those were also named.

Howseen is explicit: answers that name a brand often link to it, so the metrics are coupled. Still, the practical read is familiar from GEO research (e.g. Aggarwal et al., KDD 2024): pages that directly answer the buyer question are more likely to enter the evidence set — and naming often follows evidence the model trusts.

That differs from ghost citations, where your URL appears without your brand in the prose — see Ghost Citations. Here, the rare win condition for SMBs is both: be the page the engine quotes and be the brand it recommends.

Finding 4: Sources skew institutional and local

Among 5,613 answers with at least one citation:

French trades and regulated services drive part of the government share — buyers ask about permits, aid, and rules. The pattern still generalizes: AI answers lean on authoritative third parties, not only your homepage.

For US categories, parallel patterns show up in earned media citation share and Reddit at bottom-of-funnel. Your listing accuracy and forum presence are part of the retrieval graph, not an SEO sidebar.

What this does not mean

It does not mean 85.5% stay invisible forever. Howseen deliberately uses the first scan as baseline; only four brands had a second scan 14+ days later — too few for before/after claims.

It does not transfer 1:1 to US enterprise brands. Sentic’s 55-brand US set shows leaders holding half or more of share of voice in several verticals. Howseen describes SMB cold start.

It does not replace prompt-level monitoring. Answers drift run to run — see AI Citation Volatility and the Pineprompt October benchmark.

Playbook: sequencing work from a zero-mention baseline

1. Fix retrieval before rewriting tone

Confirm OAI-SearchBot, Google-Extended (if relevant), and other AI crawlers can reach answer pages — including CDN edge rules that override robots.txt — per Is Your robots.txt Hurting AI Visibility?.

2. One canonical page per buyer question

Match Howseen’s buyer-question framing: problem → solution → provider prompts, without stuffing the brand into the query. Lead with the answer in the first screen; product and comparison templates outperform blog fluff for citations in multiple 2026 corpora — see Product Pages vs Blog Posts.

3. Show up on surfaces AI already cites

Directories, reviews, video, and government or industry references are not “off-brand marketing.” They are evidence nodes the model sees before it ever names you.

4. Measure naming and citation separately per engine

A rising citation count with flat mention rate is a ghost citation problem, not a visibility win. Track both on a fixed prompt set — methodology in How to Measure AI Visibility Without Fooling Yourself.

How Obsurfable fits

Howseen’s study is a useful baseline statistic for SMBs. Obsurfable records buyer-style prompts and answers in a public corpus — browse in Explorer, match to your company in the App, and report mention rate, citation share, and competitors over repeated runs instead of a one-time “first scan” snapshot.

FAQ

Is 11.3% “good” for ChatGPT?

It is the share of brands named at least once in Howseen’s first-scan cohort — not your brand’s mention rate. Your denominator is answers to your buyer questions.

Should we optimize for Gemini or ChatGPT first?

Run the same prompt set on both. If one engine cites your domain and the other only names competitors, your bottleneck is engine-specific retrieval, not generic “GEO copy.”

Does own-site citation cause naming?

Not proven here — but 89.5% co-occurrence means teams should prioritize citable answer pages early, then measure whether naming follows.

Why French SMBs?

Howseen’s user base skews French; treat magnitudes as directional for other markets until you replicate with local prompts.

Bottom line

On first scan, 85.5% of brands in Howseen’s pooled data were never named; 10.6% earned an own-site citation. ChatGPT names slightly more brands than Gemini. Citation and naming move together when you win at all. Start with retrieval, canonical answers, and third-party evidence — then measure per engine, not one blended score.