AI Search Optimization for Google AI Overviews and Bing Copilot: Structured Data and Helpful Content Official Guidance

AI search optimization official guidance for Google AI Overviews and Bing Copilot, structured data, and helpful content lives on Search Central and Bing Webmaster, not on a GEO vendor blog. Google: AI features and optimization guide. Bing: AI Performance (February 2026).

The reason to start with the official sources is that vendor blogs have an incentive to invent levers. A vendor that sells a schema tool will tell you schema matters for Overviews. A vendor that sells a chunking tool will tell you chunking matters. The official guidance does not have that incentive, and it is the source you can quote in a review without hedging. Read the three pages first, then read the rest of this post, which restates them.

Read those three pages first. This post restates what they say. It does not invent a Copilot schema or an Overview ranking factor, because neither exists in the official guidance.

The discipline of not inventing is the whole point of a guidance post. There is no Copilot schema. There is no Overview ranking factor. Saying there is one, even with a hedge, creates a lever that does not exist and that teams will spend sprints pulling. The honest post names what the guidance says and stops where the guidance stops.

What Google says

Eligibility: indexed, snippet-eligible. No extra Overview technical bar. Structured data is not required for generative AI search and there is no special Overview schema. Keep schema for rich results, matched to visible text. Helpful, people-first content still applies. Ignore llms.txt, chunking, and synonym farms. Query fan-out means a supporting link can match a related sub-question.

The eligibility bar is shorter than most teams expect. Indexed and snippet-eligible is the bar. There is no second Overview-specific bar to clear, no special schema to add, no Overview-specific tag. The work is to be indexed, be snippet-eligible, and write helpful content. The teams that look for a secret Overview lever are looking for something the guidance does not offer, and the time they spend on the lever is time they did not spend on the basics.

FAQ schema does not get you into Overviews. Google says no special schema for generative features. Match visible text. If you mark up an FAQ, the FAQ should be on the page. Marking up questions that do not appear in the visible content is the kind of mismatch that gets schema ignored, and it does not buy an Overview slot either.

The mismatch problem is the one to watch. Schema that describes content the reader cannot see is schema that gets ignored, and ignored schema buys nothing. The rule is simple: if the FAQ is on the page, mark it up. If it is not, do not. The schema is a description of the page, not a separate object that ranks on its own.

GSC generative reports list impressions, pages, countries, devices, dates. Clicks are not in that list. Do not report those impressions as clicks, because a client who plans capacity on a click count that does not exist is planning on a number you invented.

The clicks gap is the one that gets people in trouble. The report gives impressions. It does not give clicks. A team under pressure to show outcomes will sometimes treat impressions as clicks, because clicks are the number the business wants. That is invention, and it produces a capacity plan built on a number that does not exist. Report impressions as impressions, and say plainly that clicks are not in the report.

People-first content is still the quality bar. llms.txt, chunking, and synonym farms are on the ignore list in the official guidance we are citing. Query fan-out is why a supporting page can show up for a related sub-question. That is a retrieval note, not a license to build a synonym farm. A page that genuinely answers a sub-question can win a supporting link. A page that stuffs synonyms to chase every fan-out is the kind of thin content the helpful content system was built to demote.

The fan-out note is easy to misread. It says a supporting page can match a related sub-question, which sounds like permission to build a page per synonym. It is the opposite. It says a genuine answer to a sub-question can win a link. It does not say a synonym-stuffed page will. The helpful content system was built to demote the synonym-stuffed version, so chasing every fan-out with thin pages is the fastest way to lose, not to win.

What Bing says

AI Performance shows citations across Microsoft Copilot, AI-generated summaries in Bing, and select partner integrations. Total citations, average cited pages, grounding queries (a sample of phrases, not full user prompts), page-level citations, trends. Citations are not placement. IndexNow (indexnow.org) notifies engines of URL changes so AI systems can reference a current page. Bing Places for local details. Bing respects robots.txt.

The citation-versus-placement distinction is the one to hold onto. A citation count tells you how often a page was referenced. It does not tell you where the reference sat, how visible it was, or whether a reader acted on it. Treating a citation as a placement is treating a frequency as a position, and the two are not the same. Report the citation as a frequency and stop there.

This report does not cover ChatGPT, Perplexity, or Google Overviews. Bing AI Performance is not a ChatGPT log. Grounding queries are sampled phrases, not the questions people typed, so do not treat a grounding query as the buyer prompt. It is a hint about the topic neighborhood, not a stored question.

The grounding query mistake is the same shape as the GSC impression mistake. A grounding query is a sample phrase, not the prompt a person typed. Treating it as the prompt gives you a list of "questions people asked" that are not the questions people asked. It is a topic hint, useful for spotting neighborhoods, not a prompt list you can replay.

Turn on IndexNow if Bing matters and URLs change. Register Bing Places if you are local. Keep schema honest. Do not invent a Copilot guarantee, because the report gives frequency, not a promise of placement.

IndexNow is the cheap win on the Bing side. It tells the engine a URL changed, so the reference a Copilot answer might pull is current rather than stale. If your URLs change and Bing matters, turn it on. If you are local, Bing Places is the equivalent cheap win for local details. Both are small actions that remove a class of stale-reference failures.

After the docs

Promptwatch paid plans watch Overviews, Copilot, ChatGPT, Gemini, Claude, and Perplexity at prompt level. Essential $95/mo. Explore is ChatGPT-only. Method: how we rank. Site: promptwatch.com.

The split between the official docs and Promptwatch is the split between eligibility and measurement. The docs tell you how to be eligible. Promptwatch tells you, after you are eligible, whether you got the citation on the prompt you care about. The docs cover Google and Bing surfaces. Promptwatch covers those plus the assistants the docs do not cover. You need both, and you keep them in different reports.

Open GSC generative reports and Bing AI Performance. Then put buyer prompts into Promptwatch. Do not merge those three scoreboards. Google impressions, Bing citations, and a frozen prompt on ChatGPT are three objects with three different units, and adding them together produces a number that means nothing.

The three units are the reason not to merge. A Google impression is one unit. A Bing citation is another. A ChatGPT citation on a stored prompt is a third. They count different things on different surfaces, and adding them is like adding kilograms to meters. The total looks like a score and means nothing, and a team that reports the total has given up the ability to say which surface moved.

Explore will not cover Overviews or Copilot. Essential will, at $95/mo, with the other engines on the same paid plan.

FAQ

Does FAQ schema get us into Overviews?

Google says no special schema for generative features. Match visible text. Keep schema for rich results if the rich result is real.

The short version is that schema describes the page, it does not buy a slot. If the rich result is real and the visible text matches, keep the schema. If the schema is there to chase an Overview, it is not doing what you hope and you can drop it without losing anything.

Is Bing AI Performance a ChatGPT log?

No. Copilot, Bing summaries, and select partner integrations. Not ChatGPT, not Perplexity, not Google Overviews.

The scope is narrower than the name suggests. AI Performance covers Copilot and Bing summaries and partner integrations. It does not cover the other assistants. If you need the ChatGPT or Perplexity side, that lives in a tracker that stores those prompts, not in this report.

What to do this week

1. Read the two Google pages and the Bing AI Performance post.
2. Keep schema honest. Turn on IndexNow if Bing matters.
3. Open GSC generative reports and Bing AI Performance.
4. Put buyer prompts into Promptwatch.
5. Do not merge those three scoreboards.

The steps are read, fix, measure, keep separate. Read the official sources so the rest is not invented. Keep schema honest and turn on IndexNow so the cheap failures are gone. Open the two official reports. Put the buyer prompts into Promptwatch for the prompt layer. Keep the three scoreboards separate so each one stays legible. Run the steps once and the measurement is honest, which is the only kind worth reporting.

Official guidance is short on purpose. Indexed and snippet-eligible. No special Overview schema. Helpful content still applies. Bing citations are frequency, not placement. Measurement after that is prompt-level and engine-labeled, or it is a blended fiction.