How to rank on ChatGPT: a practical GEO playbook

When ChatGPT answers a question that needs current information, it runs a retrieval step, pulls candidate sources, and synthesizes a response. Your material competes to be retrieved, then competes again to be the part of the retrieved set that the model finds worth quoting. Both steps can fail independently, and they fail for different reasons.

Failing the first step is usually technical or structural: the crawler cannot reach the page, the substance loads only in the browser, or the page does not clearly concern the question. Failing the second step is usually editorial: the passage is vague, promotional, undated, or makes claims a careful writer would not repeat without a source.

Diagnosing which step is failing saves months. If you are never retrieved for any phrasing of a question, treat it as an access and relevance problem. If you are retrieved but consistently paraphrased without attribution while a competitor is named, treat it as a credibility and specificity problem.

Check robots.txt for the crawlers that AI products use, and decide deliberately rather than by inheritance. Many sites carry blanket rules written for a different era, and a single inherited disallow can remove a brand from an entire category of answers. If you want the traffic and the attribution, the corresponding crawler has to be allowed.

Then confirm the page is legible without executing JavaScript. Fetch the URL and read the returned HTML. If the headings, the answer, and the key facts are absent from that response and appear only after hydration, assume a retrieval system may see far less than a browser does. Server-rendered substance is the safer default for anything you want quoted.

Finally, check the boring infrastructure: correct status codes, canonical URLs that resolve without a redirect chain, an accurate sitemap, and reasonable response times under load. None of this is glamorous and all of it is load-bearing.

A quotable passage states its claim early and stands on its own. If understanding a sentence requires three paragraphs of preceding context, it is a poor candidate for extraction. Put the direct answer immediately under the heading that promises it, then use the following paragraphs for nuance, caveats, and detail.

Match the heading to the question as a person would phrase it. Retrieval systems work with the words users actually use, and an inventive heading is a barrier rather than a differentiator. Plain and descriptive beats clever here, consistently.

Give the model something specific to quote. A named number with a source, a dated policy, a concrete threshold, a step count, or a defined term is far more liftable than a paragraph of confident generality. This is the most common gap in otherwise competent marketing content: it reads well and contains nothing extractable.

Resist the temptation to fabricate specificity. Invented statistics and unverifiable claims are the fastest way to be contradicted by a source that did the work, and a contradicted source is one that stops being quoted.

The uncomfortable part of this discipline is that a meaningful share of what a model knows about your company comes from pages you do not control: independent reviews, documentation, forum threads, comparison articles, directories, and news coverage. When those sources describe your product inaccurately or not at all, your own pages have to fight that gap rather than build on it.

The work is ordinary and slow. Make sure your product documentation is public and accurate. Correct factual errors in listings and profiles where you legitimately can. Give independent writers enough concrete material to describe you correctly. Participate honestly where your users already discuss the category.

What does not work is manufacturing this corroboration. Coordinated fake reviews and synthetic mentions are both a policy problem and a strategic one, since the pattern is detectable and the resulting description is brittle.

Build a prompt set of forty to eighty questions that reflect how buyers actually ask about your category, including problem-framed questions that do not mention your product. Run them on a fixed schedule, in a consistent way, and log three things: whether you appeared, whether you were cited with a link, and who else appeared.

The metric is a rate over the whole set, tracked across weeks. Individual answers vary between sessions, so a single result proves nothing in either direction. When the rate moves meaningfully, look at what changed on your side and on the sources that describe you.

Tie it to something commercial, carefully. Referral traffic from AI products, branded search trend, and a how did you hear about us field on your forms each capture a fragment. Report the fragments honestly rather than assembling them into a precision the data does not support.