Answer engine optimization, or AEO, is the work of getting a site quoted in AI answers: Google's AI Overviews, ChatGPT search, Perplexity and the rest. Generative engine optimization, GEO, is another name for the same thing. For Google it's mostly ordinary SEO, by Google's own account. What changes is how you measure it, because a page's citations and its position disagree often enough that one number can't stand in for the other.
What Google said in May 2026
On 15 May 2026 Google published a guide to optimizing for its generative AI features, and it settled most of the argument for Google Search. "The best practices for SEO continue to be relevant because our generative AI features on Google Search are rooted in our core Search ranking and quality systems."
The same guide says new machine-readable files, AI-specific markup and Markdown copies don't help a site appear there, because Google Search doesn't use them. It still recommends ordinary structured data, since that decides eligibility for rich results. We read every sentence of it as being about Google Search, and nothing wider. Assistants outside Google have their own crawlers and their own habits.
Citations and positions disagree
In September 2026 our tracker started recording, for every term, how many of the sources quoted in Google's AI Overview were ours, in a column beside the position. The two columns kept telling different stories.
On 9 September, two of the five sources quoted for "phoenix ai development company" were pages from our own site. One of them was a city page that didn't rank organically for that search at all. A report built on positions alone would have called that page dead while the AI Overview was quoting it.
Around the same time a new Databricks city page for Atlanta was quoted in an AI Overview four days after it was published, while it still sat on the second page of results. Exact-match tools showed no search volume for its term. If we'd judged that page by position and volume, we'd have cut a page that answer engines were already reading.
How we read citations
The rules below each come from a reading that went wrong once.
Fetch the overview properly. Google loads AI Overviews asynchronously. Our data provider returns an empty stub unless the request asks it to wait for the overview, and an empty stub reads as "not cited" when it means "not fetched". The extra request costs about six tenths of a cent per term.
Read local searches from the place. On 5 October a national reading told us a Phoenix page had fallen far down the results for a Phoenix search. Read from Phoenix and signed out, it was near the top and the first source quoted in the AI Overview. Searches with a city in them are now read with that city as the location, and a results page shorter than sixty entries counts as unreadable.
Sign out, or let the tool read. A site you visit every day is promoted in your own results. One of our city pages read page one signed in and page two signed out on the same afternoon.
Treat an error as an error. Our provider fails some deep queries and still reports success. In one September run, about a third of the zero-volume city terms came back with server errors or partial results. A blank there is unknown. It isn't a lost ranking.
Wait three weeks. A reading taken sooner measures the old page.
What ChatGPT and Bing need instead
Bing is where we learned not to trust a single source of readings. On 10 September 2026 our data provider's Bing endpoint returned Italian insurance sites for a Phoenix software search and Wikipedia for a Phoenix AI search, while the Google endpoint on the same account was fine. A position read from results like those is noise.
So Bing gets read in Bing Webmaster Tools, which is free and imports a Search Console property. ChatGPT's search has drawn on Bing's index, which makes Bing Webmaster Tools the closest measure we have for it too. Every deploy pushes its changed URLs to Bing through IndexNow, so a new page doesn't wait for a crawl.
Where an audit told us to spend effort
On 9 September 2026 we audited the AI Overviews for all 31 terms we tracked. Three findings changed what we work on.
Vendor product names are a ceiling. For unity catalog, medallion architecture, lakebase and databricks apps, every citation went to databricks.com, docs.databricks.com or learn.microsoft.com. A consultancy won't displace a vendor's own documentation in an answer about the vendor's own product. The organic position is still worth chasing. The citation isn't.
Some answers cite nobody authoritative. mosaic ai and snowflake vs databricks, at about 1,900 searches a month each, were cited from Reddit, two YouTube videos and small consultancies. Those were the openings. Our Databricks Field Guide had added a chapter on each the day before the audit.
Some expensive terms have no AI Overview at all. Five of the 31 had none, including "ai development company", which costs about $89 a click to advertise on. AEO work on those is wasted. They're won or lost on ordinary ranking.
YouTube was cited 35 times across the 31 terms, more than any other domain and more than twice the next one. A company with no video is missing the format answer engines quote most.
What we stopped doing
We don't build AEO-specific machinery beyond a few cheap surfaces: an answer paragraph under every heading, Markdown copies of each page, an llms.txt, and robots.txt rules that let the answer engines in. Everything else is the work SEO always asked for, done properly.
That means titles written for the search result, with the searcher's exact phrase first. Our event pages went through three titles in October 2026 before they led with the event's own name, which is what the ranking pages used. It also means depth, evidence a reader can check, internal links, and an ask on the page.
The measuring rules above, the citation column and the refusal to save a reading inside 21 days of the last one are built into Heddle, which we use to run search for our own products and for clients.
Questions
Is AEO different from SEO?
For Google, mostly not. Google's May 2026 guide says its AI features rest on the same ranking and quality systems as Search. What differs is measurement, because a page can be quoted in an AI Overview without ranking for the search, and the reverse.
What's the difference between AEO and GEO?
They're two names for the same work. Answer engine optimization and generative engine optimization both mean getting a site quoted in AI-generated answers. In US search volume on 10 October 2026, "aeo" came first and "generative engine optimization" second.
How do you measure AI Overview citations?
Read the AI Overview for each tracked term, wait for it to load, and count how many of its quoted sources are yours, in a column kept apart from position. Read local terms from the city, signed out, and treat errors and short results as unreadable.
How do you track whether ChatGPT cites your site?
We use Bing Webmaster Tools, since ChatGPT's search has drawn on Bing's index, and push every changed URL to Bing through IndexNow. Third-party tools that sample prompts give a second view, and we read them as a sample rather than a census.