Does passing technical SEO checks get me cited by AI?

Not on its own. citepulse.ai passes every technical readability check (robots.txt, schema, llms.txt, crawlability, sitemap) and was cited zero times across 351 AI answers about its own category. Readability is necessary but not sufficient: AI assembles answers from many corroborating third-party sources, so a technically perfect page nobody else references has nothing for the model to corroborate.

You Can't Buy AI Citations (We Tested It on Ourselves)

Q: Does an llms.txt file improve my AI citations?

No measured effect. Large studies found the vast majority of published llms.txt files receive no model requests, and no effect on citations that holds up statistically has been shown. Treat it as agent-readiness hygiene, not a citation lever. Our own site publishes llms.txt and was still cited zero times.

Q: How often should I measure AI visibility?

Repeatedly, on a fixed cadence. A single measurement is one observation from a noisy distribution: repeated runs at temperature zero still change 9 to 28 percent of decisions, and daily source overlap between runs sits around 0.34 to 0.42 where 1.0 is identical (Kirsten et al. and Schulte et al., 2026, via the Martinez survey, arXiv:2607.14035). Measure once and a citation might be luck; measure daily and a result that holds is real.

We build a tool that measures whether AI assistants cite your business. In September we pointed it at our own site. The result was uncomfortable, and it is the clearest way I know to explain why you cannot buy your way into AI answers. This first ran as a LinkedIn article, where the discussion is still open.

What our own site scored

Between September 5 and 9, 2026, we asked five AI engines the same six buying questions about AI-visibility tools, three times a day. The engines: Perplexity, Gemini, ChatGPT, Claude and Google AI Mode. That is 351 measured answers. citepulse.ai was named or cited zero times.

The engines cited 1,628 URLs in those answers. Not one was on our domain. They cited semrush.com 39 times, reddit.com 39, otterly.ai 27, llmpulse.ai 26, therankmasters.com 25. Competitors and community sites carried the answer in our own category. We were absent from it. The full run, every question, engine, date and cited URL, is published at citepulse.ai/research/citepulse-ai.

Readable is not the same as citable

Our site passes every technical readability check we run for clients. robots.txt allows GPTBot and PerplexityBot. Schema.org markup is in place. llms.txt is published. The pages render without JavaScript. HTTPS and a valid sitemap. None of it changed the result.

Readability was never the problem. Being the only source of our own story was. If you are counting on llms.txt to move citations, the measured reality is that it does not. We publish the file, we detect it for clients, and we refuse to score it as a citation factor. The full reasoning is in llms.txt: what it is, who reads it, and whether you need one. Treat it as agent-readiness hygiene, not a lever.

Why this happens: what the research says

In July 2026, Olivier Martinez published the first critical survey of generative engine optimization. It covers 45 studies from November 2023 to July 2026, each graded by evidentiary weight (arXiv:2607.14035). The conclusion is uncomfortable for anyone selling GEO as a formula.

Some interventions can change how an already-retrieved page is cited. But no reviewed technique showed a stable, cross-platform causal effect on whether a page gets retrieved in the first place, the survey concludes (Martinez, 2026). The “40% visibility gain” that shows up in GEO sales decks describes a relative maximum on one metric in one fixed-context configuration. It does not mean 40% more traffic, and it does not mean 40% more discovery.

Some tactics made things worse. One end-to-end test in the survey (SAGEO Arena, Kim et al., 2026, via Martinez) ran across 171,000 documents and 2,700 queries. Rewriting page bodies to sound more quotable cut top-10 presence after reranking by 16% and lowered final citation by 6%. In another benchmark (C-SEO Bench, Puerto et al., 2025, in the same survey), only 3 of 54 tactic-by-domain combinations showed a positive effect that held up statistically, and none worked for question answering.

AI answers are not stable

The same survey (Martinez, 2026, compiling Kirsten et al. and Schulte et al., 2026) documents something every practitioner runs into: AI answers move. Repeated runs at temperature zero still change 9 to 28% of decisions. Daily source overlap between runs sits at 0.34 to 0.42 on a scale where 1.0 is identical. In one configuration, 57.8% of ChatGPT repetitions did not trigger web search at all.

A single AI-visibility measurement is not a ranking. It is one observation from a noisy distribution. We go deeper on that in how much AI answers change day to day.

How AI assembles an answer

When a model answers a buying question, it does not pick the best-optimized page and quote it. It assembles a defensible answer from many corroborating sources: product pages, third-party reviews, Reddit threads, comparison articles, YouTube videos, press coverage. It looks for a consistent narrative it can reconstruct, not a single page it can trust.

If the only place your brand's story exists is your own domain, the model has nothing to corroborate. That is what happened to us. We had built the best page in our own category and forgotten to be talked about anywhere else.

What this changes

AI citations are not purchased. They are earned, measured, and maintained. You do not buy them by adding a trick to a page. You earn them by making one useful association repeat often enough, across credible sources, that the model can use it with confidence. Five things follow from that:

Measure across engines and repeated runs. One prompt, one engine, one time is noise. Track citation rate per engine over time, with frozen question sets, and compare changes against the engine's own day-to-day movement.
Map the sources teaching the model about your category. If Reddit, comparison articles and vendor pages are where the model learns, that is where your brand needs to appear, not just on your own domain.
Fix owned-domain clarity. Make sure your pages are readable, structured and crawlable. Necessary, not sufficient.
Build off-domain corroboration. Reviews, directories, community presence, press, creator content. The same story, repeated across sources the model trusts.
Track noise rather than pretending one result is truth. If you measured once and got cited, you might have gotten lucky. If you measure daily and the result holds, you have something real.

Measure your own citation share

Run a free audit across ChatGPT, Perplexity, Gemini and Google AI Overviews. Real buyer questions, deterministic detection, your citation share per engine, in about 30 seconds. No card, no signup.

Run a free audit →

The discussion

This study first went out on LinkedIn: the full article, and the post where the conversation is happening. Read them and weigh in there.

FAQ: buying versus earning AI citations

Can you buy AI citations?

No. No engine publishes its source weights, citations shift day to day, and no reviewed technique has shown a stable, cross-platform effect on whether a page gets retrieved. You cannot buy a guaranteed citation. You earn presence across the sources AI already trusts, then measure whether your citation share moves.

Does passing technical SEO checks get me cited?

Not on its own. citepulse.ai passes every technical readability check and was cited zero times across 351 answers about its own category. Readability is necessary, not sufficient. AI assembles answers from many corroborating third-party sources, so a perfect page nobody else references has nothing to corroborate.

Does llms.txt improve my AI citations?

No measured effect. Most published llms.txt files receive no model requests, and no effect on citations that holds up statistically has been shown. Treat it as agent-readiness hygiene. Our own site publishes it and was still cited zero times.

How often should I measure AI visibility?

Repeatedly, on a fixed cadence. Repeated runs at temperature zero still change 9 to 28% of decisions (Kirsten et al., 2026, via the Martinez survey). Measure once and a citation might be luck. Measure daily and a result that holds is real.

You Can't Buy AI Citations