What a One-Query AI Audit Is Allowed to Decide

A single prompt in a single engine is allowed to decide almost nothing about a startup's spend. It can tell me what one answer looked like on one day. Researchers at the University of St. Gallen found that 65% of cited sources turn over from one day to the next (St. Gallen GEO study, 2026). I would treat a founder who checks the company name once, sees a mention, and calls the category covered, and a founder who sees a blank and calls the category closed, as two readings of the same thin sample.

The method I would follow, with the operator's checklist and the time box, is How to Audit Your Brand's AI Visibility in 30 Minutes. That page is the deeper resource, the method. This essay is how I would use it as a founder: which steps are allowed to move a budget, which steps only earn another week of measurement, and what result would make a one-day check sufficient.

What one engine is allowed to tell me

Engines do not cite the same way, so a Perplexity hit and a ChatGPT miss are two observations, not a strategy. The GEO-16 framework analyzed 1,702 citations across three engines and found that pages cited by more than one engine scored 71% higher on quality than single-engine citations (GEO-16, arXiv). I would write "cross-engine" on the audit brief before I wrote the name of a favorite assistant.

The Citation Selection to Absorption study of 21,143 citations describes the shape of three of those engines (arXiv, April 2026):

• ChatGPT cites fewer sources and absorbs them more deeply, with a 0.27 average influence score.
• Perplexity cites the most sources and absorbs them shallowly, with a 0.06 average influence score.
• Google AI Mode and AI Overviews cite broadly, with moderate absorption.
• Gemini weights the Knowledge Graph and favors entity clarity.
• Claude is selective and uses a long context.

I would run at least five buyer prompts on each engine I care about. The prompts would be the questions a buyer asks while deciding: a "best of" for the segment, a how-to-evaluate question, and a comparison against the alternative a prospect already named. A prompt that asks the model to describe the startup back to us is a vanity check, and I would keep it out of the sample that is allowed to move money.

Why I would repeat the run

St. Gallen ran identical prompts across sessions and found brand-level visibility fluctuating at a Jaccard similarity of 0.45 to 0.59 between consecutive days. Source-level stability was 0.34 to 0.42. Their practical bar is at least seven runs per prompt per day before a brand detection rate is reliable (St. Gallen GEO study, 2026).

I would not pretend a startup will run that full grid every week from a standing start. I would set a floor that still produces a distribution: the top ten buyer prompts, three engines, three runs each, on three different days. That is 270 observations. It is enough to see a pattern, and it is the smallest sample I would let into a budget conversation. Anything thinner can set the next week's queries. It cannot set the quarter's spend.

The temptation is to want a visible-or-not answer. The studies above say the honest output is a rate. I would report the rate with the number of runs beside it, and I would leave a blank rate unpublished.

Where I would look for the mention

On-page structure has a measured threshold, and it is the smaller half of the audit. GEO-16 found that pages at a quality threshold of 0.70 or higher, with 12 or more pillar hits, reach a 78% cross-engine citation rate (GEO-16, arXiv). The same researchers added the caveat I would print under that percentage: even high-quality pages may not be cited if they live only on vendor blogs.

The larger half is where the mention sits. Muck Rack's Generative Pulse, on 25 million links, found that 82% to 89% of AI citations come from earned media in third-party publications (Muck Rack, GlobeNewswire, May 2026). I would split every audit finding into those two questions. If the pages are structured and the third-party record is empty, more schema is the wrong next invoice. The repair is a mention in a publication the engines already cite.

Citation and absorption are separate columns

Being selected as a source and shaping the answer are different events. The April 2026 paper separates selection from absorption. High-absorption pages are longer, more modular, aligned with the query, and dense with definitions, numbers, comparisons, and steps. Pages built only as question-and-answer blocks showed no absorption advantage (arXiv, April 2026).

On each key page I would look for a direct answer in the first 60 words, a named statistic with its source, one table or other structured element, and evidence a reader can check. A URL in a Perplexity footnote that contributes nothing to the paragraph is a citation without influence. I would count both, in separate columns, and I would not let a footnote total stand in for influence.

The competitive number, and when it is stable

Forrester reported that 69% of B2B marketers list AI visibility as a top CEO or CMO priority for 2026 (Forrester, March 2026). The priority explains the urgency. It does not pick the metric. The metric I would use is how often the brand is cited relative to competitors on the queries that match a live deal. St. Gallen's guidance for a stable read is a rolling window of two to four weeks, so the rate reflects sustained visibility rather than a single day's draw.

I would map about twenty buyer queries, track competitor citations across engines, and publish the map with the window attached. A startup that dominates one prompt and is absent on the prompts that precede a purchase has a source problem on those prompts. The competitor with the citation may have a weaker product and a clearer third-party record. The audit's job is to show that difference with counts.

The on-page tools in this market mostly score schema, headings, and wording. GEO-16's own recommendation is the dual strategy: on-page quality, and coverage on authoritative third-party domains. I would hire or buy against the half the audit says is empty. The sequence I keep for that work is the one on Machine Relations. The metric I would put in the memo, distinct from a raw mention count, is share of citation.

Step	What enters the memo	The smallest sample I would accept
Several engines	Presence on ChatGPT, Perplexity, Gemini, AI Mode, and Claude	Five or more buyer prompts per engine
Repeated runs	A detection rate, with the runs shown	Seven or more runs per prompt where the decision is large; 270 observations as the startup floor
Source record	The publications the engines retrieved	A list of third-party URLs, separate from our own pages
Absorption	Whether our sentences entered the answer	Evidence density, scored per page
Share against competitors	Standing on the queries that match deals	A two- to four-week rollup

The decision

I would keep the audit's output in the same shape every time I show it, so a later week can be compared with this one. Presence by engine, a detection rate with the run count beside it, the third-party URLs, an absorption note per important page, and a competitive share on a two- to four-week window. A screenshot of one answer fails that shape. A vendor score without a denominator fails it too.

I would run the 270-observation floor before I changed a content budget, a PR budget, or a tool budget. I would let a single query pick the next prompt to add. I would let the two-to-four-week rollup pick a quarterly priority. If the pages clear the GEO-16 structure bar and the third-party column is empty, the next dollar goes to a publication the engines already cite, not to another template.

Scope: this is for a startup deciding whether it is present in buyer answers, and for the first audit that will be shown to a leadership team. It uses the engines and the studies named above. It does not authorize a claim about revenue, and it does not transfer a retail or publisher finding onto a startup that has not run its own prompts.

What would change my mind: I would accept a single-day check as decisive if a later measurement, on the same prompt set, showed source turnover far below the 65% St. Gallen reported and day-to-day Jaccard similarity near the top of the scale instead of 0.45 to 0.59. I would also drop the seven-run bar if the researchers revised it with a disclosed method. I would not drop it because a dashboard rendered a confident percentage from one run.

FAQ

How often should a startup repeat this audit?

I would run a full competitive pass each quarter and keep a weekly look at the share of citation on the live prompt set. St. Gallen's day-to-day turnover, 65% of sources and a brand-level Jaccard of 0.45 to 0.59, is why a window shorter than two to four weeks stays out of the strategy memo. The weekly look is for surprises. The quarterly pass is for the budget.

Which engine should a B2B startup pick?

I would keep all five in the sample. GEO-16 found cross-engine citations scoring 71% higher on quality across 1,702 citations. A startup that tunes one engine and ignores the others is optimizing a slice. ChatGPT's 0.27 influence score and Perplexity's 0.06 describe different kinds of presence, so a win on the engine that cites widely can still be a shallow win.

What would I fix first if the audit is weak?

I would fix the third-party record first when our pages already meet the structure checks. Muck Rack's range, 82% to 89% of citations from earned media across 25 million links, is the reason. Schema and headings still matter: GEO-16's 0.70 quality threshold and 12 pillar hits line up with a 78% cross-engine citation rate, with the caveat that vendor-blog-only pages can miss anyway. Structure without an outside mention is the case I would not try to solve with more structure.

What is a citation worth if the answer never uses the page?

Less than the footnote count suggests. The April 2026 study of 21,143 citations separates selection from absorption, and question-and-answer formatting alone had no absorption advantage. I would keep a column for "cited" and a column for "absorbed," and I would let the second column guide the rewrite of a page that is already being retrieved.