Run the Same AI Prompt 20 Times to Find Your Citation Gap

An owner asks ChatGPT which company they should hire for their service in their city. A competitor gets named. The owner panics, forwards the screenshot to their marketing team, and spends the next month reacting to a single data point that would not have reproduced if they had asked again ten minutes later.

Search Engine Land's analysis of 14,472 AI citations found that business websites dominate Gemini's local answers, but repeated identical searches rarely returned the same set of sources. The prompt does not change. The answer does. That instability is the first thing to understand before you draw any conclusion about who is winning inside AI answers, because a single spot check is closer to a coin flip than a measurement.

The fix is not a better tool. It is sampling. Run the same buyer prompts enough times to see which competitor domains keep coming back, then work out what those pages contain that yours does not.

One Run Is Anecdote, Twenty Runs Is Data

AI answer engines assemble responses from a shifting pool of candidate sources. Promptwatch data tracking Reddit's ChatGPT citation share showed it collapsing 86 percent inside four days. If a domain that large can swing that hard that fast, your local competitor appearing once means very little on its own.

What matters is recurrence. A competitor cited in 2 of 20 sampled runs is noise. A competitor cited in 14 of 20 runs across three different engines is a pattern, and patterns are reverse-engineerable. Somewhere on those pages is a specific piece of evidence the model keeps reaching for because it answers the question directly.

Search Engine Journal's work on decision coverage makes the same point from the other direction: exclusion from AI recommendations is usually caused by missing evidence rather than weak authority. Models are not ranking your brand on prestige. They are looking for a page that states a price range, a coverage area, an eligibility rule, or a turnaround time in language they can lift with confidence. If nobody on your site has written that down, you cannot be cited for it, no matter how strong your domain is.

So the job is not to feel bad about a screenshot. The job is to build a sample, find the recurring sources, and identify the specific claims doing the work.

The Sampling Workflow, Start to Finish

This takes an afternoon. You need a spreadsheet, three browser tabs, and honesty about what buyers actually type.

Step one: write 8 to 10 real buyer prompts

Not keywords. Full sentences a person would say out loud. Cover five intent types:

• Service intent: "Who should I hire for emergency roof repair in Leeds?"
• Comparison intent: "What is the difference between a fixed price and hourly conveyancing quote?"
• Price intent: "How much does a full electrical rewire cost for a three bedroom house?"
• Eligibility intent: "Do I qualify for a boiler grant if I rent my property?"
• Proximity intent: "Best commercial cleaning companies near me that handle food premises"

Step two: run each prompt 2 to 3 times across three engines

ChatGPT, Gemini, and Perplexity. Fresh chat every time, no memory carryover, no follow-up questions. Ten prompts run twice across three engines gives you 60 observations, which is enough to separate signal from noise.

Step three: log four fields per citation

1. Domain cited
2. Exact page URL cited
3. Page type: service page, pricing page, comparison page, blog post, FAQ, directory listing, review platform
4. The specific claim being cited: the sentence or figure the model actually pulled

That fourth column is the one people skip and the one that carries all the value. "Competitor cited" tells you nothing you can act on. "Competitor cited for a table showing average install cost by property size" tells you exactly what to build.

For multi-location and enterprise teams, run the same prompt set per market rather than nationally. Citation behaviour diverges sharply by city, and a brand-level average hides the three regions where a regional operator is beating you on every price prompt.

Reading the Log: Citation Share and Evidence Gaps

Two outputs come out of the spreadsheet. Build both.

The citation share table counts appearances by domain across your total sampled runs. Sort descending. You will typically see three tiers: one or two competitors appearing in more than half of runs, a middle band of directories and review platforms, and a long tail of one-off mentions you can ignore. Your own row is the number that matters. Zero of 60 is a starting line, not a verdict.

The evidence gap list is built from the claims column. Group every cited claim into categories:

• Published price ranges: actual figures, bands, or "from" pricing with what changes the number
• Service coverage tables: which postcodes, regions, or property types are served, stated as a list rather than a paragraph
• Named credentialed authors: a real person with a qualification attached to the page, not "the team"
• Warranty and turnaround terms: response windows, guarantee lengths, lead times
• Eligibility and qualification rules: who can and cannot use the service, and under what conditions
• Third-party corroboration: accreditations, register memberships, independent review counts

Now compare each category against your own site. Not against what you believe is on your site. Open the pages and check. Most owners discover that they hold every one of these facts operationally and have published none of them, because the information lives in quotes, contracts, and phone calls rather than on a page a crawler can read.

Concentration matters here. Conversion and comparison intent are where these evidence gaps cluster, and those are precisely the query types AI answers pull from when a buyer is close to a decision.

Building the Response Without Copying Anyone

The instinct after an audit like this is to open the competitor's pricing page and mirror it. Do not. Copying produces a weaker duplicate of a page that already owns the citation, and it hands the model no reason to prefer you.

The correct move is to answer the same buyer question using your own operating facts. Every business already holds the raw material:

• Your last 50 invoices give you a defensible price range and the variables that move it
• Your job scheduling data gives you real average turnaround times by service type
• Your coverage map is a fact, not a marketing claim, and belongs in a table
• Your qualified staff have real credentials that can be attached to the pages they are competent to write
• Your terms document already contains warranty periods nobody has published in plain language

Map each recurring gap to one page. If price intent prompts keep citing a competitor's cost guide, publish your own cost page built from your invoice history, including the honest ranges and the reasons a job lands at the top of the band. If eligibility prompts keep citing a competitor's FAQ, write the eligibility rules you actually apply, including who you turn away.

Then link. New evidence pages sitting in isolation get crawled slowly and cited rarely. Add contextual internal links from your existing ranking service pages into each new evidence page, using the language of the buyer prompt rather than the page title. A structured Autopilot SEO Engine approach treats those links as part of the build, not a follow-up task, because the whole point is making the new fact reachable.

The Citation Gap Checklist

• 8 to 10 buyer prompts written in customer language, covering service, comparison, price, eligibility, and proximity intent
• Each prompt run 2 to 3 times in ChatGPT, Gemini, and Perplexity, fresh session every time
• Every citation logged with domain, URL, page type, and the exact claim being cited
• Citation share table sorted by frequency, with your own appearance count recorded
• Evidence gap list grouped into the six proof categories
• Each recurring gap mapped to one page you can publish honestly from internal data
• Named author with real credentials attached to every evidence page
• Internal links added from existing ranking pages into each new evidence page
• Entity facts (name, address, service list, coverage) consistent and crawlable across the site
• Re-run date scheduled in the calendar before the audit closes

What to Fix First

Rank by frequency, not by effort. Work in this order:

1. The single most-cited claim category across your sample. If price ranges appear in 60 percent of cited pages and you publish none, that is the first page. One page, built properly, beats five thin ones.
2. Comparison prompts where a competitor is cited and you are absent entirely. These are buyers in active evaluation. A comparison page written from your own methodology, honestly stating who each option suits, is defensible and rarely duplicated.
3. Author attribution on pages you already have. This is the cheapest fix on the list. Real names, real credentials, real bios on existing service and guide pages, done in a day.
4. Coverage and turnaround facts in structured form. Convert paragraphs into tables and lists. Same information, dramatically more extractable.
5. Entity consistency. Mismatched business details across your own pages undermine everything above, because a model that cannot reconcile who you are will not risk recommending you.

Re-run the full sample monthly. Same prompts, same engines, same log. The measurement that matters is movement in citation share: from zero mentions in 60 sampled runs to appearing in 11 of 60, and the ability to name which evidence block caused it. That is a claim you can defend to a board or a bank manager.

Discovery is now an operating discipline rather than a campaign. Prompts get sampled, gaps get logged, evidence gets published, and the cycle repeats. If you want a baseline before you start, run the audit against your own domain first and see what an engine can actually find. Get a Free Audit and start the log with real numbers rather than a screenshot.