Getting cited by AI search: why ChatGPT is not quoting you

What is GEO, and how does it differ from SEO?

SEO gets a page into a ranking; GEO gets a passage into a quote. In classic search the user sees ten blue links and decides which to click. In a generative engine the engine writes the answer, picks the sources itself, and the user often never clicks at all. The place you compete is not the results list but the inside of the answer.

This does not mean SEO is over — both engines crawl the same page and technical health matters for each. What changed is that a well-written page is no longer enough. Even if the engine likes your page, it moves on when it cannot find a passage it can quote.

How does ChatGPT read a web page?

Not as a whole. It splits the text into passages, scores each passage against the query separately, and takes only the highest scorer into the answer. The practical consequence: what decides the outcome is not your page's average quality but the quality of its single strongest passage. A good ten-paragraph article can lose to an ordinary page that has one quotable paragraph.

The second consequence of that reading is harsher: every sentence beginning "as described above" or "thanks to this approach" becomes unquotable, because lifted out of context it no longer says what it meant. Most corporate copy is written exactly that way.

Why is AI not showing our website?

• The page carries promotion rather than an answer: "the sector's leading solution partner" corresponds to no query at all.
• The answer is gathered at the end of the page; finding nothing in the early passages, the engine leaves.
• The numbers have no source. Rather than quote a claim it cannot verify, an engine skips it — a wrong number costs more than no number.
• Headings are topic names: had the heading been the question the user typed, the match would be direct.
• No date. Freshness is a trust signal; an undated page reads as stale and loses to a dated rival.

How do I optimise my site for AI search?

Should I allow AI bots in robots.txt?

In classic SEO, robots.txt says what must not be crawled. In GEO its function inverts: an engine can only quote your page if its own crawler can reach it. The trap here is that a blanket `User-agent: *` line does not cover every engine.

Google's AI side looks for a line written for `Google-Extended`, and Apple's for `Applebot-Extended`. Without those lines your site keeps appearing in ordinary Google search but can drop out of Gemini and Apple Intelligence answers — and nothing warns you that it happened. That is precisely why this site's robots.txt lists 36 crawlers by name.

Do I need an llms.txt file?

Not required, but cheap. robots.txt tells a crawler what it may reach; llms.txt tells a model what the site contains — a map showing in one read which page answers which question. Because a hand-written llms.txt starts lying at the first content change, it has to be derived from the content; on this site the file contains not one hard-coded link and is generated entirely from the content.

How do I track traffic coming from ChatGPT?

AI traffic does not arrive as its own channel in analytics; you have to look at the referring hostname: chatgpt.com, perplexity.ai, claude.ai, gemini.google.com, copilot.microsoft.com. ChatGPT also appends `utm_source=chatgpt.com` to citation links, so some of the web-side traffic arrives already tagged. Collecting those five hostnames into one channel group is the first step of measurement.

But measurement has a limit here, and anyone unaware of it draws the wrong conclusion: mobile apps and in-app browsers frequently drop the referrer when opening an external link. That visit shows up in analytics as direct traffic. Your AI traffic is therefore always under-reported; the number you see is a floor, not the real value.

Does appearing in AI answers bring clicks?

Because the engine already answered the question, the user often never reaches the source. Measured per click that looks like a loss. But the few clicks that do arrive are different in kind: the user has compared the options, has seen you as a candidate, and comes on that basis. Where a visitor from classic search opens with "who are these people", a visitor from here usually opens with "do you do this particular thing".

The right reading: treat AI visibility as a brand line rather than a traffic line. "How many engines mention us when I ask this question" is more informative than "how many clicks came in" — and the first one you can measure by hand, every month, in fifteen minutes.

Is it easier to rank in AI search in a smaller language?

Because structured Turkish sources are scarce. On an English query the engine chooses among dozens of well-written pages; on the Turkish version of the same query it usually meets undated, unsourced content copied from itself. That makes becoming the primary national source in your own field far cheaper than it would be in English.

The window is not permanent. As Turkish content gets structured this advantage closes; a good page written today is worth more than the same page written two years from now.

How did Neuros apply this on its own site?

• Every service, product, industry and case page carries a 40–75 word answer capsule that names its subject.
• Body headings are questions rather than topics, in the form the user would type into a search box.
• Every number is tied either to a verified primary source or to a verifiable fact of our own — no unsourced statistics are written.
• robots.txt allows 36 crawlers by name, including `Google-Extended` and `Applebot-Extended`.
• /llms.txt and /llms-full.txt are derived from the content and never edited by hand.
• An /api/mcp endpoint lets agents ask the site directly instead of fetching and interpreting pages.

Do I need to blog more to appear in AI search?

What decides it is not the count but the number of questions answered. A site publishing four general posts a month loses to one page answering a single topic across fifteen distinct questions. A generative engine scores passages; fifteen questions means fifteen passages. Before raising your publishing rate, count how many distinct questions your existing pages answer.

Our competitor appears and we do not — why?

Three common reasons, all technical. First, access: with no line for that engine's crawler in robots.txt your page is never read. Second, form: an answer buried mid-paragraph cannot be extracted as a passage. Third, verifiability: rather than quote an unsourced claim the engine skips it and takes your competitor's sourced sentence instead.

AI is saying something wrong about us — what can I do?

You cannot edit a model's memory directly, but in engines that fetch live you can fix the source. Wrong information usually comes from an old page of yours, a stale directory or a third-party listing. Publishing the correct answer on your own site, plainly, dated and sourced, is what lets the engine find the right passage on its next crawl.

Should I write separate content for AI?

No — write the same content differently. Producing a separate "for AI" version creates duplicate content and doubles maintenance. What wins is building one page readable by both: question headings, short context-free answers, tables and sources. All three make it faster for humans too.

“In a generative engine the place you compete is not the results list but the inside of the answer.”