Generative engine optimization works in four steps, and each one is a place to lose. An AI engine retrieves a handful of pages for a question, reads them, decides which it trusts, and writes an answer that quotes a few. If your page is not retrieved, not readable, not trusted or not quotable, you are not in the answer. GEO is the work of fixing each of those four, in that order. This post walks through what happens inside the engine and what you can change on your side.
If you want the overview first, start with our guide to generative engine optimization. This post is the mechanism.
Step 1: Retrieval. The engine goes and looks.
People imagine that ChatGPT answers from memory. For anything current, local or specific, it does not. ChatGPT search, Perplexity, Copilot and Google’s AI Overviews all start with a live retrieval step: the question is turned into one or more searches, the engine pulls a short list of pages, and only those pages are read.
That short list is the whole game. Google’s AI Overviews draw on Google’s own index and ranking. Copilot draws on Bing. ChatGPT and Perplexity run their own crawlers and lean on third-party search as well. In every case the pages that already rank for the plain search are the pages most likely to be retrieved. This is why GEO sits on top of SEO rather than replacing it: a page on the third page of Google is rarely in the handful an engine reads.
Retrieval also fails for a dumber reason. Each engine has a crawler with its own name: GPTBot and OAI-SearchBot for OpenAI, ClaudeBot and Claude-SearchBot for Anthropic, PerplexityBot for Perplexity, Google-Extended for Google’s AI training. A robots.txt rule, a security plugin or a firewall setting that blocks them means your pages are never fetched. We see this on a surprising share of the sites we scan. One nuance: blocking Google-Extended does not keep you out of AI Overviews, which use the normal Googlebot crawl.
What to do: allow the AI crawlers in robots.txt and at the firewall, make sure the page content is in the HTML rather than loaded by JavaScript after the fact, and keep working on ordinary rankings. Our free GEO scan checks crawler access in about a minute.
Step 2: Reading. The engine extracts what the page says.
Once a page is retrieved, the engine does not read it the way a person does. It breaks the page into passages and looks for the ones that answer the question directly. A passage that says “A fractional CMO is a senior marketing executive who works with a company part time” is usable. A passage that says “In today’s fast-moving landscape, businesses face many challenges” is not, and the engine moves on.
Structure helps here more than people expect. Headings that are real questions, a definition in the first sentence under each, short paragraphs, lists where the content is a list, and tables where the content is a comparison all make passages easier to pull out cleanly. The research paper that coined the term GEO, published in 2023 by a team from Princeton and Georgia Tech, found that adding statistics, quotations and cited sources to a page raised its visibility in generative answers by up to 40% in their benchmark, while keyword stuffing did nothing.
What to do: answer the title question in the first 60 words of every page and post. Use specific numbers, names and dates. Put one idea in each paragraph. Mark up what the page is with schema (Organization, Service, FAQPage, Article, LocalBusiness) so the facts are machine-readable without guessing.
Step 3: Trust. The engine decides whether to believe you.
Engines are built to avoid confidently repeating something false, so they weigh agreement. If your site says you are a dermatology practice in Chattanooga, the engine checks whether your Google Business Profile, your LinkedIn page, the directories, the local press and the review sites say the same thing. When they do, the claim is treated as established. When they disagree, or when your site is the only source, the claim is treated as unverified and the engine prefers a competitor whose facts line up.
This is the part of GEO that looks least like marketing and matters most. Old addresses on directory listings, a different phone number on Yelp, a description on Clutch that still mentions a service you dropped, and a category on Google that does not match the one on your site all cost you citations.
Third-party mentions carry more weight than your own pages. A client naming you in an interview, a podcast appearance, a local news story, a case study on a partner’s site, an award listing: each is an independent source an engine can cite instead of you, and engines like citing sources that other sources cite.
What to do: one name, address, phone number, description and category everywhere. Fix every listing you can find. Then earn mentions from sources that are not you: reviews, interviews, press, client sites.
Step 4: Writing. The engine composes the answer and picks who to name.
The final step is generation. The model writes the answer from the passages it trusted, and attaches citations to the sources it used. Different engines show this differently. Perplexity shows numbered sources next to each claim. ChatGPT search shows a list of linked sources under the answer. Google’s AI Overview shows link cards beside the text. Copilot shows footnotes. In each case a reader sees your name only if your passage made it through the first three steps and was the clearest statement of the point.
Two things decide whether you are named rather than merely used. The first is whether your passage reads as a complete, standalone statement that the model can paraphrase without stitching it to another source. The second is whether your business is named inside the passage: “Spectruss, a marketing and advertising company in Chattanooga, has worked with more than 250 companies since 2012” can be cited as a fact about Spectruss. “We have worked with hundreds of companies” cannot, because the model does not know who “we” is once the passage leaves the page.
What to do: write the facts about your business in the third person at least once per page, with the name attached. Keep an entity paragraph (who you are, where you are, what you do, since when) on the homepage, the about page and the footer, worded the same each time.
How this differs from how Google ranks a page
Google ranks a whole page for a query and gives it a position. A generative engine retrieves passages, scores them for the specific question, and can cite a paragraph from a page that would never rank first in a classic search. That cuts both ways: a well-written answer on a modest site can be cited over a vague page on a big one, and a page that ranks first can be skipped if its content does not answer the question directly. The full comparison is in GEO vs SEO: what is the difference.
How to tell it is working
Measure it the way the engines behave. Pick the ten questions a buyer would ask before hiring you, ask each engine those questions once a month, and record whether you are named, whether you are cited, and who else is. Google Search Console now reports impressions from AI features in Search, and Bing Webmaster Tools reports citations from Copilot, so two of the engines give you first-party numbers. The rest you check by hand. When the mention count climbs before traffic does, that is normal; a share of the value is the reader who never clicks but already knows your name when they call.
Where to start
- Run the free GEO scan for crawler access and the basics.
- Read how to optimize for generative search for the step-by-step work.
- If you would rather someone did it, our SEO and GEO work covers both kinds of visibility, and it starts with a 15-minute call.
