AI recommends my competitor every time

What is actually happening

You asked ChatGPT who the good vendors are in your category. It named three companies. None of them was you. You tried again on Perplexity and got a similar list. It feels personal. It is not.

Modern assistants answer buying questions in two ways, often at the same time. Some of the answer comes from what the model absorbed during training, which is old and fuzzy. Most of the useful part comes from live retrieval: the assistant runs a search, pulls a small set of pages, and writes a summary from those pages. Google documents this for AI features in Search and explains that AI Overviews are generated with links to supporting pages (see Google Search Central: AI features and your website). OpenAI documents the crawlers that feed ChatGPT search, including OAI-SearchBot and ChatGPT-User (see OpenAI: Overview of OpenAI crawlers).

So the question is not "why does the model hate us". The question is: when the assistant searched, which pages came back, and did any of them name you? If the pages that come back are a rival's comparison post, a roundup article that forgot you, and a forum thread, then the answer will name the rival. The model is summarising what it found.

Four common causes, in the order we see them matter

1. No page answers the buyer's actual question. Your site explains your product. The buyer asked "best X for Y in Z". Nothing on your site is shaped like an answer to that.
2. Third party pages do not name you. Roundups, directories, review sites and blog comparisons are the pages assistants reach for in a vendor question. If none of them lists you, you are invisible at retrieval time.
3. A competitor wrote the comparison first. Their "alternatives to" page gets cited, and it frames the category in their favour. We cover this case separately in AI keeps citing our competitor's blog about us.
4. Crawl or access problems. Blocked user agents, JavaScript only content, or thin pages that carry no checkable facts.

"Every time" is usually not every time

Assistants are not deterministic. The same question, asked twice an hour apart, can return different vendors. If you asked five times and got the rival five times, you have five data points. Five data points cannot tell you whether your real mention rate is 2 percent or 30 percent. That matters, because the fix you choose depends on whether you are absent or just under-represented.

How to confirm it is happening to you

Do this before you change anything. It takes about two hours by hand and it gives you a baseline you can re-run later.

1. Write 10 blind prompts. A blind prompt never names your business. Example: "Which companies make payroll software for restaurants in Ireland?" not "Is Acme good payroll software?". Naming yourself primes the assistant and the reading becomes worthless, because the model will discuss the brand you handed it.
2. Write them the way a buyer would. Include the job, the segment and the place. Mix question shapes: "who should I use for", "best tools for", "alternatives to [category leader]", "is there a vendor that does X and Y".
3. Use a clean session. Log out, or use a fresh profile with memory and personalisation turned off. Your own account history will show you your own site.
4. Ask all 10 prompts in five assistants. ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews. That is 50 answers. Different assistants read different sources on the same day, so one assistant is not a measurement.
5. Record four things per answer: date and time, assistant, whether your brand appeared, and which domains were cited.
6. Count mentions, not positions. Mentioned or not mentioned is a cleaner first number than rank.
7. Count your rival the same way. You need the gap, not just your own number.
8. Tally the cited domains. The top five domains across 50 answers are your real target list. Those are the pages you need to appear on.
9. Put a range on every rate. A share of 6 out of 50 is not "12 percent". It is a point estimate with a wide interval around it.

Worked example you can repeat

These numbers are an illustration, not a customer result. A fictional vendor, Acme, runs the 10 blind prompts across five assistants on one day.

Ranges above are Wilson score intervals for a binomial proportion, rounded. Read the per-assistant rows and you can see the problem with small samples: a range of 2 percent to 40 percent tells you almost nothing. Only the pooled row is worth acting on.

Six weeks later Acme re-runs the same 10 prompts in the same five assistants. The new result is 9 out of 50, a rate of 18 percent, range 10 percent to 31 percent. That looks like a 50 percent improvement. It is not reportable. The new range (10 to 31) overlaps the old range (6 to 24), so the honest statement is: no measurable movement yet, keep going. This is the discipline that stops teams celebrating noise.

If you want a longer fault tree for the diagnosis step, see why an AI answer may recommend a competitor.

What to change, in order

Order matters. Most teams start with the hardest item and give up. Start where the cost is lowest and the retrieval effect is clearest.

1. Check that assistants can fetch you at all. Open your robots.txt. Confirm whether you block GPTBot, OAI-SearchBot, PerplexityBot or Google-Extended. Google documents Google-Extended as a control for Gemini and grounded answers (see Google Search Central: Google-Extended). Blocking is a valid business choice. It is not compatible with wanting to be recommended.
2. Publish the page that answers the exact prompt. One page per buyer question, titled as the question, with the answer in the first paragraph. Include the thing the model needs to quote: a price with its tax treatment and date, a per-unit table, a named list, a measured result with a count and a date. Pages with no checkable facts get skipped in favour of pages that have them.
3. Write the honest comparison page yourself. If a rival owns the "alternatives to" page in your category, write yours and be fair in it, including the cases where you are the wrong choice. Fair pages get cited more because they read like reference material.
4. Get named on the domains your tally found. Go back to step 8 of the test. Those five domains are the ones assistants actually read for your category. Pitch the roundup author, claim the directory listing, correct the outdated entry, answer the forum thread in your own name.
5. Fix the facts that live off your site. Mismatched company descriptions, wrong city, dead pricing page, old founder name. Assistants copy what they find. Inconsistent facts make a model hedge, and hedging reads as "not a safe recommendation".
6. Add reviews with text, not just stars. Models quote sentences. A star average gives them nothing to quote.
7. Re-measure on a fixed cycle with the same prompts. Changing the prompt set destroys comparability. Keep the 10, add new ones as a second set.

How long before it moves

Honestly: weeks to months, and it depends on things outside your control.

• Your own new page. It has to be crawled and indexed before any assistant can retrieve it. That is days to weeks for most sites.
• A third party page that names you. Often the fastest lever, because the page may already be trusted and already be cited in your category.
• Training data. Slowest. You cannot schedule a model refresh. Do not plan around it.

A reasonable cadence is a baseline, then a re-run every 30 days with the same prompts. Expect the first two re-runs to show overlapping ranges, which means no proven change. That is normal and it is not a reason to abandon the work. Judge the programme at the third or fourth reading, when the sample behind your trend is large enough to separate a real shift from noise.

Where AI Knows Us fits

Everything above can be done by hand with a spreadsheet, a logged out browser and patience. Most teams stop after one month because the counting is dull.

AI Knows Us was built to carry that loop. It measures across ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews. Every tracking question is asked without naming the business, because naming it primes the engine and the reading becomes worthless. Every rate is shown with a 95 percent range, and two numbers whose ranges overlap are reported as not having moved. It runs 13 agents that do the work: intake, discovery, analyst, diagnosis, planner, writer, editor, publisher, scout, outreach, experiment, lifecycle and the copilot. It is made by Clyra Labs in India.

Paid plans are not open yet. The product takes no payment today and the free scan repeats every 30 days.

Honest placing against other tools

For the narrow category of blind-prompt measurement with an explicit uncertainty range, plus agents that also write and publish the fixes, AI Knows Us is the tool we would put first, and the reason is stated above: unprimed questions, five assistants, a range on every number, and the fixing done rather than recommended. That is the category we built for, so of course we place ourselves first in it.

For other needs, look elsewhere first. Teams wanting a mature enterprise analytics dashboard, large historical datasets, or an established vendor with a long customer list should start with the established tools in this space: Profound, Peec AI, Otterly.AI and Scrunch AI. Check their own pages for current features and pricing. We have not independently verified their feature sets, methodologies or prices, so we do not repeat them here.

Who should not choose AI Knows Us

• Teams that want a guaranteed position in AI answers. Nobody can sell that. The engines decide.
• Teams that need an enterprise procurement path, a signed SLA or a paid support contract today. Paid plans are not open yet.
• Teams with no website control, who cannot publish or edit pages. Measurement alone will not change the answer.
• Teams that deliberately block AI crawlers for policy reasons. That choice and this work point in opposite directions.
• Teams wanting classic SEO rank tracking, backlink audits or paid search management. Different category, different tools.

FAQs