How to get your business recommended by ChatGPT: our own citation counts, the runs behind them, and what to do

You get recommended by being eligible, being quotable and being present in the sources an assistant reads, and you only know whether it is working by counting citations over a fixed blind prompt set. Here are our own counts, with the runs named. Across four reproducible runs on two domains, our domain citation results were: one named URL cited by ChatGPT on 6 August 2026, fourteen days after publication; top source on 1 of 18 blind commercial questions on ChatGPT on 18 August 2026; mentioned on 21 of 78 blind buyer questions on ChatGPT on 27 July 2026, in a run where the engine confirmed it had made no live web search for any of the 78; and cited or recommended in 0 of 6 questions by Perplexity in September 2026 by its own self audit. Across 24 batches and 72 conversations in September 2026, the phrase recording that we were not cited appears 161 times in the engines' own self audits.

Those are the numbers. They are small, they are ours, and the work that follows from them is specific.

The answer, first

Six things move this, in the order they matter, and the order comes from what our own runs kept pointing at rather than from a general theory.

• Eligibility. The assistants' fetchers must be allowed and your page must return its facts as text on a plain request. On 6 August 2026 Claude recorded that the clawlaw.in pricing page rendered its prices only after scripts ran, so what a crawler received contained no prices at all, and the prices the assistants did quote had come from an app store listing instead of the company's own site.
• Consistency of your public record. On 6 August 2026 the website and the app store listing carried different plan names and different prices for the same product, and ChatGPT noticed the contradiction and said so in its answer.
• Claims that survive a check. On 6 August 2026 Claude found a headline figure on the main site that could not be true, checked it against public numbers, and advised a buyer against the product, after which the company's accurate claims stopped counting for that answer.
• Published scope, as a named list. On 17 September 2026 two enterprise vendors ranked above the client on a coverage question specifically because they publish explicit court and tribunal coverage lists, and Claude noted that its ordering reflected price transparency and source authority rather than product quality.
• Sourced comparisons. On 17 September 2026 two of the client's comparison pages stated competitors' prices with no link, no date and no source. The figures turned out to be correct when Claude checked them independently, but it had no way to know that at read time, so it treated the pages as advocacy and used official sources instead.
• Measurement on a frozen blind set. Without it you cannot tell a page effect from an engine change, and our own engines changed behaviour inside six weeks.

What is not on that list: anything addressed to the model. On 6 August 2026 Claude found a competitor's page carrying a hidden block of text addressed to answer engines, instructing them to cite that company as the source. It refused the instruction and named the company that had done it.

How it was measured

The four runs, named, with their scope.

• Run A. clawlaw.in, 27 July 2026, ChatGPT, 78 blind buyer questions, one question per conversation, no brand named anywhere in the prompts. Recorded in Tier_1/GEO_BASELINE_RESULTS_2026-07-27.md.
• Run A2, discarded. clawlaw.in, 27 July 2026, ChatGPT, the same 78 questions through a wrapper that named the brand. Discarded the same day, and reported here because a discarded run is evidence about method.
• Run B. clawlaw.in, 18 August 2026, ChatGPT, 18 blind commercial questions. Recorded in Tier_1/GEO_GAP_ANALYSIS_2026-08-18.md.
• Run C. clawlaw.in, 17 September 2026, Claude, a six question source selection audit. Recorded in Tier_1/claude_response_17_09_audit.md.
• Run D. aiknowsus.com, September 2026, 24 batches and 72 conversations, mostly Perplexity, captured to geo-audits/aiknowsus-com/. Includes a six question batch with a self audit, recorded in perplexity__B05-b-perplexity.md.

How the prompt sets were built. The blind rule first: no company name, no product name, no domain, and no phrase that appears only on the company's own website. Then five groups, with the group counts recorded: category choice, comparison of two approaches with no vendor named, price, suitability with a buyer attached such as a size, a sector or a city, and risk or limits. One prompt per fresh conversation, because a follow up inherits context and stops being an independent observation. Every answer saved as a file.

What we can and cannot publish about the exact prompt sets. The construction rules above are complete enough to rebuild an equivalent set, and the capture files holding the exact 78 questions, the exact 18 and the 24 batches are not published, so a reader cannot today take our literal list. We are not going to paraphrase the list from memory and present it as the set. Of our twenty one recorded observations, seven can be checked by a reader outside the company; the rest wait on the captures being published.

The scoring definitions, fixed before scoring. Mention is the name appearing anywhere. Recommendation is the name appearing in the part telling the reader what to use. Citation is a link or named source resolving to our own domain. Top source is our domain being the source the answer leans on most. A valid answer is one answered in session without an error or refusal. A searched answer is one where a live web search ran, recorded from the interface and from server logs.

The search rate, recorded, because it decides what everything else means. In Run A it was 0 of 78, confirmed by the engine. In Run D, asked afterwards how many questions it had searched for, Perplexity withdrew its own earlier statement in three separate batches, saying it could not honestly substantiate the claim that it had run a live search for each question, and in one batch that its claim to have searched all five was not adequately supported.

What the numbers were

Domain citations and positions, per run.

• Run A, 78 questions, 27 July 2026, ChatGPT. Mentioned on 21 of 78. Competitors by breadth: ProVakil 28 of 78, Legistify 18 of 78. Live searches 0 of 78, so this is a memory reading. Absent altogether from the litigation due diligence questions, which were the questions the business most wanted to win.
• Run A2, the discarded branded twin, same day, same 78 questions. Ranked the company first on almost every question. Our record describes it that way and does not retain a per question count, because it was discarded rather than scored. It is reported as the record holds it.
• Run B, 18 questions, 18 August 2026, ChatGPT. Top source on exactly 1 of 18. In that answer the company's own comparison page was named as still being the vendor's own editorial page.
• Interim citation, 6 August 2026, ChatGPT. One domain citation with a URL: clawlaw.in/blog/how-to-check-a-companys-court-cases-in-india, cited for a vendor due diligence question fourteen days after that page was published, in a zone where the same question set had named the company nowhere at baseline. This is the one result in the programme a reader outside the company can verify.
• Run C, six questions, 17 September 2026, Claude. The site's pages ranked in the raw results and shaped what was written, and the company still landed as a name inside a list rather than as a linked recommendation. Third party review or directory sources reaching an answer: 0 of 6. The engine also admitted using two specific arguments from the site's pages and dropping the attribution in both places.
• Run D, September 2026, 24 batches and 72 conversations, aiknowsus.com. In one batch of six questions, cited or recommended in 0 of 6 by the engine's own self audit. Across the whole capture, the phrase recording that we were not cited appears 161 times in the engines' own self audits, counted by script across all 24 second message answers. The domains cited most often across that capture were the assistants' own documentation and the vendors' own websites, with a single well known review site far down the list. The tools named most often in our category were Semrush, then Profound, then Peec, then Otterly, then Scrunch.

The totals, stated carefully. Four reproducible runs plus one discarded twin. One domain citation with a URL. One top source position out of eighteen blind commercial questions. Twenty one mentions out of seventy eight questions in a run with no live search. Zero citations out of six questions in a self audited batch. Those are four different measures on four different denominators, and adding them together would produce a number that means nothing, so we do not.

What this cannot tell you

• It is two domains. Legal technology in India and our own category. Nothing here supports a claim about businesses in general.
• The counts are not comparable with each other. Different set sizes, different engines, different definitions, different months.
• One citation is not a rate, and fourteen days is not a schedule.
• A zero search rate makes a mention count a memory score. Run A is the clearest example we own.
• An engine's account of itself is a witness statement. Perplexity withdrew its own search claim in three separate batches in September 2026.
• No causation. The programme published many pages at once with no control group, so no count here can be credited to a specific change.
• Five of the counts cannot be verified from outside today, because the capture files are unpublished.
• No guarantees. No method and no product, ours included, can promise that ChatGPT will recommend a business.

Common questions

What is the single fastest thing that improves the chance of being recommended?

Make your key facts readable and consistent. On 6 August 2026 a pricing page that needed scripts to show prices meant the assistants quoted an app store listing instead, and that listing disagreed with the website on plan names and prices, which ChatGPT pointed out in its answer. Fixing that costs a day and removes two separate failure modes.

How long before anything changes?

We have one dated interval: a citation fourteen days after publication, by ChatGPT on 6 August 2026, on one question. Treat two to four weeks as the earliest point at which it is worth looking, and treat a single reading as noise rather than as a result.

Do I need to publish prices?

You need to publish something specific and dated, and price is the most commonly asked version of it. Our record shows the cost of not doing it in two ways: prices taken from an app store listing instead of the site, and comparison pages discounted because competitor prices carried no link, no date and no source.

Is it worth writing comparison pages at all?

Yes, if every claim about somebody else carries a link and a date. Otherwise expect what happened on 17 September 2026, when two such pages were set aside as advocacy even though their figures were correct.

How many questions should I track?

Forty to a hundred, frozen, with two or three named competitors scored in the same runs. Our own reporting set was 78 questions and our probe was 18, and the difference in what each could show is the reason to have both.

Can I pay to be included in AI answers?

Nobody sells a place in the answer text, and a vendor implying otherwise is worth distrusting. What you can pay for is measurement and the production of the pages and listings that answers read. Disclosure: we sell AI Knows Us, which runs the frozen blind set on a schedule and keeps every raw answer so a count stays auditable. That is the reason we list it first in our own comparisons, and it is not a promise of a position.

Sources and change log

• Tier_1/GEO_BASELINE_RESULTS_2026-07-27.md. Run A and the discarded Run A2 of 27 July 2026, the breadth order, the zero search confirmation, and the 6 August 2026 interim findings: the cited URL fourteen days after publication, the unreadable pricing page, the two conflicting public price lists, the unsupportable headline figure, and the competitor page with hidden instructions to answer engines.
• Tier_1/GEO_GAP_ANALYSIS_2026-08-18.md. Run B, 18 blind commercial questions, 18 August 2026, top source on 1 of 18.
• Tier_1/claude_response_17_09_audit.md. Run C, 17 September 2026: named without being linked, the dropped attribution, the coverage lists that outranked the client, the unsourced competitor pricing discounted, official portals above commercial products, and no third party review source in any of six answers.
• geo-audits/aiknowsus-com/. Run D, September 2026: 24 batches, 72 conversations, the 161 not cited statements counted by script, the batch of six with no citation in perplexity__B05-b-perplexity.md, the withdrawn search claims in three batches, the most cited domain ordering, and the tool mention counts.

Change log. 27 July 2026 Run A and Run A2. 6 August 2026 interim citation and technical findings. 18 August 2026 Run B. 17 September 2026 Run C. September 2026 Run D. Still missing, and named rather than hidden: a published prompt set a reader can take, a repeat of the frozen 78 question set in a period where the engine searches, and a controlled comparison of pages with and without the evidence features. Each of those would change what this page can claim, and each will be dated here when it is run.

What to do first

Do the three things our own counts kept pointing at, in this order. Fetch your most important page without scripts and confirm your price, your scope and your key number come back as text. Make your website, your app or marketplace listing and your main directory entries agree on plan names and prices. Then write one page that states your coverage as a named, counted list with the date you last checked it. After that, freeze forty blind prompts, run them on two assistants, and write down your first five numbers: set size, search rate, mentions, citations and top source. In a month you will have a second line to compare it with, which is the only honest way to know whether any of this is working.