Why an AI answer may recommend a competitor: a diagnostic checklist and a repeatable test

An AI answer usually names a competitor for one of four reasons: the answer never searched and is reciting what the model remembers, your pages are not readable or not eligible, the competitor publishes a checkable fact that you do not, or your own prompt was doing the work. In our own blind run of 78 buyer questions on 27 July 2026, ChatGPT named ProVakil on 28 questions, CLAW on 21 and Legistify on 18, and confirmed it had run no live web search for any of the 78. So in that run the competitor was ahead on breadth by seven questions, and no page on anybody's website was involved in the result at all.

That is the most common finding in this work and the least expected one. Before you rewrite anything, find out whether the answer was reading the web or reading its own memory.

First separate a one-off answer from a repeatable pattern

A single answer is not evidence. The same question asked twice in two fresh sessions can name different companies on the same day. So the first step is arithmetic, not diagnosis.

Ask the same question five times, each in a new conversation, and count how many of the five named the competitor and how many named you. If the competitor appears in five of five and you in zero of five, you have a pattern worth investigating. If it is three of five against two of five, you have noise, and the next step is a larger question set rather than a theory.

This is also the point to be honest about what an engine's own summary is worth. In batch after batch during our September 2026 audit, Perplexity declined to name a competitor as winning most often, saying its own previous answers had not produced a comparable live tested result to support such a claim. That is the clearest statement we have that one run does not establish a ranking, and it came from the engine itself.

Check technical eligibility

Six checks, in order, each of which has produced a real problem in our own programme.

• Robots file. Read your own file and confirm the assistants' fetchers are not blocked, including by a wildcard rule somebody added for scrapers.
• Server response to a plain fetch. Request your key page without running scripts and read what comes back. 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.
• Facts as text, not as images. A registration number, a price or a coverage list inside a picture is invisible to a reader that only has the text.
• Consistency across 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 that a headline figure on the main site could not be true, checked it against public numbers, and advised a buyer against the product. After that, the company's accurate claims stopped counting for that answer.
• Access logs. Look for the assistants' fetchers in your own logs around the time of a run. This is your only first party evidence that a page was read at all.

Check whether the answer used web search and inspect its citations

This is the check that changes the diagnosis most often, and it takes two minutes.

Look for the search indicator in the interface. Then look at the citations: count them, and write down each domain. Then ask the engine, in a second message, how many of the questions in the run it searched for. Keep that reply as data rather than as fact. In September 2026, asked exactly that, Perplexity withdrew its own earlier statement, saying it could not honestly substantiate the claim that it had run a live search for each question, and in another batch that its claim to have searched all five was not adequately supported. That happened in three separate batches.

Three outcomes and what each one means:

• No search. The answer is model memory. Nothing you publish this week will change it this week. On 27 July 2026 ChatGPT confirmed it had run no live web search for any of 78 questions, which is what made the breadth order a ranking of what it had absorbed rather than of what was true that week.
• Searched, competitor cited, you absent. A source selection problem. Read the cited pages and find the checkable thing they publish.
• Searched, you cited, competitor recommended. Your page was used as background while another source decided the recommendation. This is the most fixable state, because you already have the engine's attention.

Check whether the prompt favours competitor attributes

Read your own prompt as a hostile reviewer would. Four ways a prompt decides its own answer.

The brand is in the question. The strongest case in our record: on 27 July 2026, two runs of the same 78 questions happened on the same day, and the one whose wrapper named the brand came back ranking it first on almost every question, while the blind one put it second by breadth and absent altogether from the litigation due diligence questions it most wanted to win. The branded run was discarded, because the only thing it had measured was our own prompt.

The question carries an attribute only one vendor has. Asking for the tool with the largest published coverage list is a question about published coverage lists.

The question is a how to question. On 17 September 2026 Claude recorded that on questions about how to look a case up, official court portals took every position above any commercial product, and said plainly that no commercial product should rank above the official portal for a question about using that portal. If your set is full of how to questions, your competitor is not beating you; the government is beating both of you, correctly.

The question has no place or no buyer in it. A category question with no country attached invites an answer built from whatever is written about anywhere. On 17 September 2026 one well known review site appeared in the raw results and was discarded, because the list it offered was of American products and so was not an answer to an India question.

Run a matched-prompt test

The full protocol, so the comparison between you and a named competitor is fair.

Pick one named competitor. Not the category. One company, so every count has a comparison.

Write 30 to 60 blind prompts across four groups, and record the group counts: category choice, direct comparison of two options without naming either, price, and suitability for a stated buyer. Freeze the list.

Run on two assistants, one prompt per fresh conversation. Record the product and mode used.

Score six fields per answer. Did it search. Every company named, in order. Whether you were named. Whether the competitor was named. Whether either domain was cited. Which domain was the top source.

Report four counts over the same denominator. Prompts naming the competitor and not you. Prompts naming you and not the competitor. Prompts naming both. Prompts naming neither. Then report the same four over the subset of prompts where the engine actually searched, because that subset is the only one that describes today's web.

Say in advance what each result means. Competitor ahead only in the no search subset points at memory and third party presence. Competitor ahead only in the searched subset points at your pages. Both ahead of you in both subsets means your category question is being answered by someone else entirely, often an official body or a directory. Neither named on most prompts means your prompts are informational, not commercial, and need rewriting.

What we have measured against this protocol. The counts we hold are from the runs above, not from a designed head to head. On 27 July 2026, over 78 blind questions on ChatGPT with zero live searches, the breadth counts were ProVakil 28, CLAW 21, Legistify 18. On 18 August 2026, over 18 blind commercial questions on ChatGPT, the company was the top source on exactly 1 of 18. We have not yet re run the matched set in a period where the engine searches every question, so we cannot publish the searched subset counts that this protocol asks for, and we are not going to estimate them.

A worked example of the diagnosis

On 17 September 2026, during the clawlaw.in programme, Claude was asked how to find every case filed against a company. It ranked two enterprise vendors above clawlaw.in, and it said why: those vendors publish explicit court and tribunal coverage lists, and its ordering reflected price transparency and source authority rather than product quality. In the same audit it recorded that two clawlaw.in comparison pages stated competitors' prices with no link, no date and no source; the figures turned out to be correct when it 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.

Read as a diagnosis, that is not a mystery. The competitor won on two publishable attributes: a named coverage list and a sourced price. Both are documents, not products. Both can be written in a week. Neither required the assistant to like anybody.

What the test cannot tell you

• It cannot tell you the internal reason. An engine's explanation of its own choice is a statement, not a mechanism, and our own capture includes an engine withdrawing such a statement three times in September 2026.
• It cannot prove your change caused the improvement. Retrieval behaviour changes on its own. A before and after with no control cannot separate the two.
• It cannot be pooled across engines or dates. Keep ChatGPT and Perplexity counts separate, and keep July counts separate from September ones.
• It cannot see a competitor's private advantage. Placement inside a directory, an old link from a trusted site, or a mention in documentation you cannot read.
• It cannot promise a position. Nothing in this method, and no product including ours, can guarantee that an assistant will name you.
• Most of our own counts are not externally checkable yet, because the capture files are unpublished. The exception is the ChatGPT citation of clawlaw.in/blog/how-to-check-a-companys-court-cases-in-india on 6 August 2026, fourteen days after that page went up.

Troubleshooting checklist and sources

Run these in order and stop at the first one that is false.

• Did the answer search at all? If not, stop treating this as a page problem.
• Is your key page readable without scripts, with its facts as text?
• Do your site, your app listing and your directory entries agree on plan names and prices?
• Does every claim on your most important page survive an independent check?
• Does the cited competitor page carry a named list, a dated price or a stated limit that yours does not?
• Is your comparison page sourced, with a link and a date next to every competitor figure?
• Is the question actually a how to question that an official portal should win?
• Does your prompt name your brand, your product or a phrase only your site uses?

Sources. Tier_1/GEO_BASELINE_RESULTS_2026-07-27.md for the 27 July 2026 blind run, the discarded branded twin, the breadth order, the zero search confirmation, and the 6 August 2026 findings on crawlable prices, the two public price lists, the unsupportable headline claim and the competitor's hidden instruction block. Tier_1/GEO_GAP_ANALYSIS_2026-08-18.md for the 18 question run of 18 August 2026. Tier_1/claude_response_17_09_audit.md for the 17 September 2026 findings on advocacy, coverage lists and official portals. geo-audits/aiknowsus-com/ for the September 2026 audit of our own domain, 24 batches and 72 conversations, including the three batches where the engine withdrew its search claim.

Common questions

A competitor appears every time I ask. Is that proof they are better optimised?

It is proof they are better remembered or better published, and those are different things. Check the search indicator first. If the answer did not search, what you are seeing is what the model absorbed months ago, and no amount of on page work changes that answer this week.

Can I ask the assistant why it chose them?

Ask, and record the reply as evidence rather than as a cause. Sometimes it is directly actionable, as on 17 September 2026 when Claude named published coverage lists and price transparency as the reason for its ordering. Sometimes it is not supportable, as in the three September 2026 batches where Perplexity withdrew its own claim to have searched.

Should I write a comparison page that beats theirs?

Write one, and source it. An unsourced comparison page is read as advocacy: on 17 September 2026 Claude discounted two such pages even though their competitor prices were correct, because at read time there was no link, no date and no source. Put the link and the date next to every figure about somebody else.

How long should I wait before measuring again?

Give a new page at least two to four weeks before you conclude anything. The one dated case we have is a citation fourteen days after publication, by ChatGPT on 6 August 2026, on one question. One case is not a schedule.

Is it worth being named without a link?

Yes, and score it separately so you can see it. On 17 September 2026 Claude recorded that clawlaw.in pages shaped what it wrote while the company landed as a name in a list rather than a linked recommendation, and in the same audit that it had used two of the site's arguments and dropped the attribution in both places.

Does paying for ads or a directory listing fix this?

A directory listing can help, because directories are among the sources these answers read, and a paid listing is still a page with your facts on it. It is not a placement in the answer, nobody sells that, and a page that claims otherwise is worth distrusting.

What to do first

Take the single question where the competitor keeps winning. Ask it five times in five fresh conversations and write down how many named them and how many named you, with the date. Check the search indicator on each. If none searched, spend the next month on consistency and third party presence. If they searched and cited the competitor, open the cited page and find the one checkable thing on it that is missing from yours, then publish that thing with a date on it.