In 65 cold scans of businesses with no relationship to us, 67.7% received no mentions at all across five engines. The answers were not empty; they named other brands.
Of 65 businesses scanned cold, 44 — 67.7% — received no mention at all. Not a low score: absent, across three prompts on five engines, fifteen chances each. The answers were not empty. They named six to eight other businesses instead.
This is our cold-market baseline. It is the closest thing we have to an unbiased picture of how the general market looks to an AI assistant, and it is materially worse than anything measured on brands that have been onboarded, prompt-engineered and tracked. It is also a sample of 65, which is a number we will return to in some detail below, because what it cannot support matters as much as what it can.
Our cold lane scans a domain a business enters itself. There is no onboarding call, no prompt engineering, no relationship with us and no opportunity for anyone to select questions the brand is likely to win. Up to three prompts are run against five engines, producing fifteen brand-and-engine cells per business. Between July and August 2026, 65 such scans completed.
Set against the corpus of tracked brands — where prompt sets are designed, and the brands themselves are larger — the level is three to five times lower on every engine.
| Engine | Cold scans, presence per run | Tracked corpus, presence per run | Ratio |
|---|---|---|---|
| Gemini | 12.3% | 40.1% | 3.3× |
| ChatGPT | 10.3% | 36.3% | 3.5× |
| Claude | 7.8% | 36.4% | 4.7× |
| Google AI Overviews | 6.7% | 19.7% | 2.9× |
| Perplexity | 5.1% | 26.1% | 5.1× |
Two details are worth more than the headline. The first is that the two columns broadly agree on engine order — Gemini is the most generous in both, and Perplexity and Google AI Overviews are the two least generous in both — despite the two samples sharing no brands, no prompts and no selection process. The agreement is not exact: Claude and ChatGPT are separated by 0.1 points in the tracked corpus and swap places against the cold column, as do Perplexity and AI Overviews at the bottom. That is a partial independent replication of how the engines behave, arrived at by accident.
The second is that the answers were full. Gemini named an average of 7.7 brands per cold answer and ChatGPT 6.5. The demand is being served. Somebody is being recommended to these businesses' customers, and it is not them. Own-domain citations appeared in only 3% to 15% of cold cells, which is the mechanism in plain sight: a business whose pages are not retrieved is not named.
The honest claim is: two-thirds of the businesses we scanned cold were absent from AI answers. The claim we are not making is that two-thirds of businesses are absent. Those are different statements and only the first is measured.
Sixty-five is a small sample. It carries a wide confidence interval on its own, and it was not drawn at random. These businesses selected themselves by entering a domain into a free scan, which almost certainly selects for firms already thinking about AI visibility — that is, firms with some marketing sophistication and some existing web presence. If that selection biases the result at all, it biases it towards visibility, which means the true market absence rate is more likely to be higher than 67.7% than lower. We say more likely, not certainly; we have not measured it.
The prompts were self-chosen, three per business, with no quality control on phrasing. A badly formed prompt can return an answer that names nobody, and we cannot separate businesses that lost the answer from businesses whose prompt produced no competitive answer to lose. And there is no industry field on these records, so we cannot yet say whether absence is evenly spread or concentrated in particular sectors — a gap we found in our own data and are fixing rather than working around.
What makes the figure worth publishing anyway is the replication described above. A directional number from a small sample is worth little on its own. A directional number from a small sample whose engine-level structure broadly reproduces a corpus twenty times larger, drawn from entirely different brands, is worth reporting with its limits attached.
The first practical consequence is that most businesses are asking the wrong first question. The question is not what our AI visibility score is. It is whether we appear at all, on the questions our customers actually ask, before any of the optimisation conversation begins. For roughly two-thirds of the businesses we have scanned cold, the answer to that prior question was no, and every subsequent metric was moot.
The second is that absence is not neutral. If an AI assistant answers a buying question by naming six to eight companies and yours is not among them, the answer has not failed. It has succeeded, for a competitor. This is a different competitive situation from a search results page, where being on page two is a poor outcome but a recoverable one. In an AI answer there is no page two — across our corpus, between 64% and 78% of all brand mentions sit at position one and almost nothing appears past position five.
The third is that the gap between cold and tracked brands has two causes and both are real. Tracked brands perform better partly because their prompt sets are engineered to be winnable, and partly because they are stronger, better-established businesses. That means the tracked figure overstates what a typical business should expect, and the cold figure understates what a well-run programme can reach. Neither number alone describes anyone's situation. The useful reading is the distance between them, which is the size of the addressable problem.
A larger cold sample is the obvious test, and the one we intend to run. Several hundred scans with an industry field attached would turn this from a directional finding into something closer to an estimate, and would show whether absence is concentrated in particular sectors. We will publish that figure whichever direction it moves.
A sample drawn without self-selection would be stronger still. If a randomly drawn set of registered businesses in a defined sector returned materially lower absence than 67.7%, the self-selection effect would become the story and this baseline would need restating.
The prompt-quality question is separable and testable. If the absence rate is being driven mainly by poorly formed self-chosen prompts rather than by genuine invisibility, then re-running the same businesses on well-formed specific question prompts should lift results towards the question-prompt band we measure elsewhere, at around 46%. We have not run that test. It is the single cheapest thing that could substantially revise this finding.
Engine behaviour is also moving underneath the measurement. After a model transition on 7 July 2026, answers in our corpus began naming roughly twice as many brands each. If that inflation continues, absence rates should fall for reasons that have nothing to do with the businesses being measured — more seats at the table, not better candidates. Any repeat of this study will need to report the two effects separately.
The cold-scan sample sits inside a wider corpus of 22,012 completed AI answer runs between 15 April and 21 August 2026 — 11,554 production runs and 10,458 staging runs — across five engines (ChatGPT, Claude, Gemini, Perplexity and Google AI Overviews), approximately 110 brands, and 5,440 captured citation links. The cold lane contributed 65 completed scans between July and August 2026, each up to three prompts across five engines, for up to 975 brand-and-engine cells.
Presence is the share of completed prompt runs in which the brand was mentioned, computed with read-only SQL against the live databases. A cold business counts as absent when no mention was recorded in any of its cells.
Mentions are resolved by entity resolution with a three-state verdict, adopted on 18 August 2026, replacing a pure word-boundary string matcher that counted any occurrence of the brand's name as a mention regardless of whether the answer concerned that company. The verdict is asymmetric by design — more willing to record uncertainty than to assert a match — so it reduces reported presence rather than inflating it.
No onboarding, no prompt engineering, no prior relationship, no curation of the prompt set by us. This is what distinguishes it from the tracked-brand corpus, whose prompts are designed per brand and therefore measure how a chosen brand performs on chosen questions rather than how the market looks.
Production and staging were kept separate throughout and never pooled. The cold-scan finding has an unusual second form of replication: the engine-level ordering it produces broadly tracks the ordering independently measured on the far larger tracked-brand corpus, which shares none of its brands or prompts.
Figures are quoted as measured. Caveats are stated before findings rather than after them, and data gaps we found in our own records — including the missing industry field on cold-scan leads — are listed rather than omitted.