GEO vs SEO: what changes when AI answers the question

SEO is the work of earning a click from a list of links. GEO — generative engine optimization — is the work of being a sentence an answer engine quotes when it answers the question itself. The same URL can do both. The decisions are not the same, and a rank report does not tell you whether a model can lift a true sentence from the page.

The engines people mean are ChatGPT and its search mode, Perplexity, Google AI Overviews, Claude, and the copilot-style boxes that sit on top of a search results page. There is no public, stable formula for which source they pick. CiteScore does not log into those products. The free check reads public files. The $39 audit scores the HTML we can fetch. This page is the comparison. A shorter definition of the category is what GEO is.

Two jobs

Classic SEO asks whether a page can be crawled, understood, and ranked for a query, then clicked. The levers practitioners already know still apply there: a page Googlebot can fetch, a title that matches the query, internal links, and the authority other sites assign you. CiteScore does not grade title-tag templates or backlink velocity. Those are a different audit.

GEO asks a narrower question. If a model has to write the answer, can it lift a definition, a product name, and a fact from your page, and can it tell which company those words belong to? The unit is a passage that still makes sense when it is pulled out of the layout. A page can rank and still be a slogan. A page can be clearly written and still be blocked.

What changes

The passage replaces the position

A results page can send a click to a URL whose answer sits halfway down, behind a hero. An answer engine quotes a short span. If the first 80 words do not say what the product is, who it is for, and what it does, the model has to guess or quote someone who did say it. Put that sentence in the HTML. A slogan can stay in the design. It cannot be the only sentence.

Write H2s as the questions a buyer asks, and answer each one in the first sentence under the heading. Pricing belongs in text, with the unit and what is included. A comparison a buyer will ask — “you vs the tool they already pay for” — belongs on your domain, in your words. If you do not write it, the quoted source is someone else’s list.

The same standard applies to evidence. A number, a date, and a method are quotable. “Industry-leading” is not. If the fact comes from someone else, link the original. The citation checklist turns this into 40 checks you can run before you pay anyone, including us.

Crawlers split into jobs

SEO’s crawler question is mostly “can Googlebot and Bingbot fetch this?” GEO adds agents that do not share a switch. GPTBot is OpenAI’s training crawler. OAI-SearchBot is the search crawler. ChatGPT-User is a fetch a person triggered. PerplexityBot and Perplexity-User are the same kind of split. Google-Extended is Google’s token for whether content helps improve Gemini Apps and Vertex AI generative APIs. Google says it does not change inclusion or ranking in Google Search.

The crawler page is the precise version of what a CiteScore audit records: GPTBot, PerplexityBot, and Google-Extended, each as allow, disallow, or unspecified. OAI-SearchBot, ChatGPT-User, ClaudeBot, and Googlebot are not in that trio. The AI Overviews row does not read Google-Extended. It reads answer-shaped copy and FAQ or entity markup. Blocking a training crawler on purpose is a policy. A leftover Disallow: / from staging is an accident. The free check shows the three tokens the audit uses. It does not query the answer engines.

AI Overviews are a Google Search feature. A page Google cannot index is not a candidate for that surface. That is the part of SEO that GEO for Google still depends on. It is not a claim that a top-10 rank guarantees the overview will quote you. We do not have a public overlap number, and we do not invent one.

Schema extracts. It does not rank.

Organization, SoftwareApplication, and FAQPage JSON-LD give a parser a labeled copy of facts that are already in the HTML. The name in the schema has to match the name in the H1. FAQPage is for questions a person can read on that page. Google stopped showing FAQ rich results in May 2026, so this markup will not restore an accordion under a blue link. Keep it for extraction. Skip it when the questions are not on the page.

llms.txt is the other file people add when they hear “AI search.” It is a curated reading list. It is not a ranking factor we have ever seen an engine publish, and a missing file does not zero a CiteScore audit. Ten live SaaS files, and the specific choice each one gets right, are on the llms.txt examples page. The files were fetched. The citations were not inferred from them.

What you can measure, and what you cannot

Search Console will tell you impressions and clicks for a query. It will not tell you the sentence ChatGPT used last Tuesday. A single answer moves with the prompt, the user’s location and memory, and whether browsing or a tool was on. Anyone selling a stable “you are #4 in ChatGPT” number should show the prompt, the settings, and the raw answer, then show them again next month. CiteScore does not sell that number.

What you can measure on the page is whether a careful reader, and therefore a model, has the raw material: a definition, one spelling of the product, a sourced fact, markup that matches the prose, and a crawler that is allowed to fetch it. Off-site mentions still matter. Answer engines lean on sources that are already cited elsewhere — docs, comparisons, directories, community threads. We do not crawl the web for those mentions. An enterprise monitor that samples prompts over time is a different product, compared on the enterprise tools page.

What does not change

• The page has to be true. A model repeats confident sentences. A statistic you would not put in a contract does not become safer because an answer engine might quote it.
• The page has to be fetchable. One canonical host, HTML in the first response, no accidental disallow of the whole site.
• The entity has to be one thing. The same spelling in the title, the H1, and the schema. If the company and the product have different names, say both.
• Classic search still sends customers. Blue links have not been replaced. GEO sits next to that work for the queries where the engine answers in prose.

Free check, then the $39 audit

Run the free check first. It looks at one host and reports whether the homepage responded, whether robots.txt allows or blocks GPTBot, PerplexityBot, and Google-Extended, whether /llms.txt is a non-empty text file, whether a sitemap is present, and which JSON-LD types appear on the homepage. A passing row means the signal is present. It is not a live answer from ChatGPT or Perplexity.

The $39 GEO audit is one public URL, a 0–100 CiteScore, and about ten prioritized fixes, as a PDF. The score comes from on-page signals — definitions, entities, evidence, schema, freshness, crawler access — and, when a language-model key is configured, a review of the HTML we fetched. We do not log into ChatGPT, Perplexity, or Google AI Overviews. The per-engine snapshot is an inference, and the report says so.

A practical order: the free check for the door, the 40-point checklist for the page, the $39 audit when you want that page scored. Weekly rescans are on the roadmap. They are not this product.