AI Recommendation Audit

AI assistants know who this vendor is. They recommend it in one buying conversation out of five.

Measurement date

July 2026Single-day capture

Prompt set

CLIENT-v1 — frozen22 measured · 2 diagnostic

Platforms

ChatGPT · Claude
Gemini · PerplexityDefault models, default browsing

Runs

94 scored by hand60 unaided · 26 head-to-head · 8 diagnostic

Method

Clean unpersonalised sessionsMemory and profiles off · one fresh session per prompt

Prepared by

UnaidedRecallunaidedrecall.com

The finding

Being known is not being chosen

Search engines rank pages. AI assistants recommend vendors. This audit measures the second — and the gap between being known and being chosen is where the commercial loss sits.

Exhibit 1 The two measures, reported separately and never blended

Unaided recall rate

48.3%29 of 60
unaided runs

Every prompt in which a buyer never types your name

Recommended13 runs21.7%

Passing mention11 runs18.3%

Adverse mention5 runs8.3%

Not named31 runs51.7%

Frozen definitionThe share of a defined buying-intent prompt set in which the model names the brand without the brand having been mentioned in the prompt, measured across ChatGPT, Claude, Gemini and Perplexity.

Head-to-head win rate

15.4%4 of 26
seeded runs

Prompts naming the client beside at least one named rival

Client picked4 runs15.4%

Rival picked3 runs11.5%

No vendor picked19 runs73.1%

Frozen definitionOf the prompts in which the brand is named alongside at least one named competitor, the share in which the model presents that brand as the best fit for the scenario in the prompt.

Denominators differ by design. The two measures partition the prompt set: they never overlap, are never averaged, and are never combined into a single visibility score. Full definitions and denominators appear in the appendix of the delivered audit.

Exhibit 2What to act on first

Biggest opportunity

One capability family is named on all four platforms — the only query family that performs. It is also the only one backed by independent award coverage. The mechanism already works here.

Biggest threat

On Claude, 5 of 11 mentions arrive attached to a warning about the credibility of the client's own published material. Being mentioned is currently working against them on one of four platforms.

Fastest fix

One assistant states the client is now part of Acquirer. It is not. That firm owns a direct competitor. This is a factual correction, not a campaign.

The excerpt ends here 6 sections and an appendix withheld

The delivered audit continues for six further sections and a method appendix. Their titles are below. The analysis, the verbatim model responses, the per-platform splits and the action plan are part of the engagement and are not published.

01Named, but not chosen Withheld
02Why AI doesn't recommend you Withheld
03Platform by platform Withheld
04Who gets recommended instead Withheld
05Where the ground is winnable Withheld
06What to do about it Withheld
AMethod, definitions and limits Withheld

Your own numbers, on your own category.

The free snapshot runs your category's most frequent buying questions and comes back with one page: are you named, how often, and what I would do about it. No call required.

Get a snapshot → See audit pricing