AIO: How Google's AI Answers Choose Their Sources, and How to Tell Whether You Are One

What is AIO?

Google says its AI Overviews now have more than 2.5 billion monthly users, and AI Mode more than a billion. Until this summer, a site owner could not see how often their own pages were linked in either. On 31 August 2026 Google finished rolling out a Search Console report that counts exactly that, and nothing more: impressions, with no clicks and no queries.

That count is the first official measure of AIO, the work of getting a site's pages used as sources, and its brand named, in AI-generated answers. People use the term two ways. Narrowly, AIO means AI Overviews optimization: earning a link in the summary Google shows above its results, which SEOs shorten to "AIO". Broadly, it means artificial intelligence optimization for any AI answer, including Google's AI Mode, ChatGPT, Gemini, Microsoft Copilot and Claude. Many teams now call the same work AEO (answer engine optimization) or GEO (generative engine optimization). AIO is the umbrella over both, and this guide uses it that way.

The labels differ in emphasis more than in substance. AEO leans on the answer itself, a direct and quotable response to a question. GEO leans on inclusion, whether a language model names and cites you when it composes a reply. AIO, used broadly, takes in both and adds the reporting. Google does not draw the line at all: its guide to generative AI search says optimizing for it is optimizing for the search experience, "and thus still SEO." Where the labels do lead to different work, our comparison of AEO, GEO and SEO sets it out.

This guide works through how Google picks its sources, what to change on a site, how to measure the result and what to do in the next 90 days. Its argument is that the doing part of AIO is mostly SEO that was always worth doing, as Google's documentation says. The measuring part is new, and it is where most of the money in this market is wasted.

The stakes are clicks. In an Ahrefs analysis of 300,000 keywords published in February 2026, the top-ranking page's desktop click-through rate was 58% lower when an AI Overview appeared, against a forecast built from keywords without one. Pew Research Center, tracking 900 US adults' browsing in March 2025, found people clicked a traditional result on 8% of visits to results pages with an AI summary, against 15% without one.

Being cited seems to soften the loss. Seer Interactive's April 2026 update, covering 5.47 million queries across 53 accounts, found cited brands had a higher organic click-through rate than uncited ones in every month of 2025. Seer adds that it cannot claim causation, because higher-authority brands are also more likely to be cited.

How do Google AI Overviews and AI Mode choose their sources?

Google's guide says AI Overviews and AI Mode rely on the same core ranking and quality systems as the rest of Search. The models ground their answers in pages retrieved from Google's index, and both may use what Google calls query fan-out: the model generates a set of related searches across subtopics, then draws supporting pages from all of their results. That is why an AI Overview often links to pages that do not rank for the query you typed.

To be eligible, a page has to be indexed and able to show a snippet, and the site has to be included in the Search generative AI setting in Search Console, which is on by default. Google's page on AI features and your website states that "there are no additional technical requirements." The guide lists what you do not need: no llms.txt or other AI text files, no chopping content into small pieces, no rewriting for every long-tail phrasing, no special schema.org markup. Seeking inauthentic mentions across the web, it adds, "isn't as helpful as it might seem."

Independent data fits the fan-out explanation. In July 2025 Ahrefs found that about 76% of URLs cited in AI Overviews also ranked in the top 10 for the same query. When it re-ran the study in March 2026 on 863,000 keyword results and 4 million cited URLs, the share was 37.9%, and 31% of cited pages ranked outside the top 100. The authors offer fan-out as a hypothesis, not a finding: citations increasingly come from the results of Google's sub-queries rather than the original one.

Where the citations land matters as much. Among the 50 domains Google's AI Overviews cited most across more than 3 million US queries in September 2026, YouTube took 22.9% of those citations and Reddit 18.5%, according to Ahrefs' monthly tally.

Ranking still matters, but the unit of competition is the sub-question, so a page that answers the follow-ups a buyer would ask has more ways in than one that answers only the head term. And the answer is assembled from more than your site, which puts video, forums, reviews and trade press inside the scope of the work.

An AIO checklist you can actually run

Each item has a test. Run the list on the 20 pages that matter most commercially before the rest of the site.

1. In Search Console, open Settings, then Search generative AI, and confirm the site is included. That is the default, but Google's help page for the control says an excluded site's content generally drops out of AI Overviews and AI Mode within one to two days.
2. Run URL Inspection on each priority page. It must be indexed, and nothing may stop a snippet: no nosnippet, no restrictive max-snippet, no data-nosnippet around the paragraph that holds the answer. Check robots.txt and your CDN's bot rules for anything that blocks Googlebot.
3. Put each answer in text at the top of its section, under a heading that names the question. Google asks for important content in text form; if an answer sits behind a click or inside an image, check the rendered HTML in URL Inspection.
4. List the five to ten follow-up questions a buyer asks after the main one (price, alternatives, how long it takes, what goes wrong) and answer each on the page or on a page it links to. That is your fan-out surface. Do not give each its own thin page; Google treats pages mass-produced to manipulate rankings as scaled content abuse.
5. Add something a competitor cannot copy from the same results page: your own prices, test results, photographs, customer data or a named expert's first-hand account. Google's guide puts "non-commodity content" first on its list.
6. Mark up what is on the page, and only that. Structured data is not required for AI features, Google says, but it must match the visible text, and it still matters for rich results. Our schema markup guide covers the types most sites need.
7. Bring Google Business Profile and Merchant Center up to date. Google names both as inputs that help products and services appear in AI responses.
8. Verify the site in Bing Webmaster Tools, whose AI Performance report shows how often Microsoft Copilot and Bing's AI answers cite your pages.
9. Check robots.txt against the search crawlers of the other assistants. OpenAI says a site that blocks OAI-SearchBot will not be shown in ChatGPT search answers. Perplexity recommends allowing PerplexityBot, its search crawler, for a site to appear in its results. Anthropic documents three user agents and says blocking Claude-SearchBot or Claude-User may reduce a site's visibility in Claude's search results. The training crawlers, GPTBot and ClaudeBot, are separate switches.
10. Earn mentions where AI answers look: a video that properly answers the buyer's question, genuine participation in the forums your buyers read, reviews on the sites your category uses, coverage in its trade press. Our guide to a Reddit strategy for Google and AI Overviews covers the forum side.

For a quick structural pass over one page, the free AIO readiness checker flags gaps in headings, schema, text depth and links, and our explainer on how an AIO score is built and how to raise it describes the weighting. Treat any readiness score as a lint check: it says a page is well formed, not whether an assistant will cite it.

How do you measure AIO?

Sort every AIO number into one of two kinds before it goes into a report: counted by the platform, or estimated from a sample. They answer different questions and should never be added together.

What Google counts

Search Console's Generative AI performance report, introduced in June 2026 and open to every site since 31 August, shows impressions from AI Overviews and AI Mode by page, country, device and date. An impression means a link to your site was shown in one of those features. There are no clicks and no queries; Google says more metrics will follow. Its chart counts by property, so two of your links in one answer make one impression, while the page table counts by page. The totals will not match, by design.

Clicks are in Search Console, but not separately. Google says traffic from AI features is reported in the Performance report under the Web search type, mixed with everything else. Its counting rules say a click on a link in an AI Overview counts as a click, and every link in an Overview is assigned the single position the Overview occupies. A page cited in an Overview at the top of the results can therefore report a better average position than its own listing holds.

What Microsoft counts

Bing Webmaster Tools' AI Performance report, in public preview since February 2026, counts citations of your pages in AI answers across Microsoft Copilot, Bing and some partner experiences, with the grounding queries behind them. In June Microsoft added Citation Share, your percentage of all citations for a grounding query, which it calls "an observational metric" and not a ranking. It does not report clicks.

What proves a change

Analytics adds sessions from assistants that pass a referrer. Clicks from Google's AI features arrive from Google like any other search visit, so analytics cannot single them out. And none of these numbers shows that your work caused a change, because Google updates its models and competitors publish all the time. To credit a change, compare the pages you changed with similar pages you left alone over the same weeks. Our guide to proving an SEO change worked with a control group walks through the method.

What an AIO tracker can and cannot tell you

For assistants that report nothing to site owners, and for the question no platform report answers (who gets named instead of you), the only instrument is a sample: send the same prompts repeatedly and count. Every AIO tracker does some version of this. The useful ones admit the result is a sample.

In research published in January 2026, SparkToro and Gumshoe.ai had 600 volunteers run 12 recommendation prompts through ChatGPT, Claude and Google's AI Overviews (or AI Mode, when no Overview appeared), 2,961 times in all. For ChatGPT and Google's AI, the chance of two runs returning the same list of brands was under 1 in 100, and the same list in the same order came up about once in 1,000 runs. SparkToro's Rand Fishkin concluded that a brand's position in an AI answer means little, while the share of answers that mention it, across dozens to hundreds of prompts each run several times, is a reasonable metric. The work is not peer reviewed, and the authors disclose that Gumshoe sells tracking software.

Here is what that means in numbers, using Harbour & Finch, a sample client. Track 20 buyer questions and run each 10 times on one assistant: 200 answers. The firm is named in 46 of them, 23%, with a 95% confidence interval of about 18% to 29%. Next month it is named in 60 of 200, 30%, with an interval of about 24% to 37%. The intervals overlap, and a test of the difference gives a z-score of about 1.6, short of the 1.96 that 95% confidence requires. A seven-point rise that looks like progress on a slide is inside the noise.

The real interval is wider still, because answers to the same prompt resemble each other: ten runs of one prompt are not ten independent observations. To get within about five points either side at a rate near 25%, you need roughly 290 independent answers per assistant per period.

A tracker has no impressions, so it cannot say how many people saw an answer. Its answers come from its own accounts and locations, a stand-in for what your buyers see. Its prompts are a set someone chose, and the numbers move when the set changes. And no tracker sees Google's internal data; Google's guide warns against tools that claim to use "internal" Google metrics.

What a well-run tracker does tell you is still worth having: which rivals the assistants name instead of you, and which sources they cite when they do, a ready-made list of places to earn a mention. On a fixed prompt set, with the margin shown, it also gives the direction of travel. For the full method, read our guide to measuring AI visibility as share of model.

An AIO strategy and 90-day roadmap

An AIO roadmap is mostly a matter of order. Eligibility comes before rewriting, because a page that cannot show a snippet cannot be cited however good it is. Rewriting comes before chasing mentions, and the measurement at the end needs a control set chosen at the start.

Weeks 1 and 2: the baseline

Export three months of the Generative AI performance report and of Web performance data. Choose the 20 pages that matter commercially, plus a matched set of similar pages you will leave alone as the control. Write 20 to 30 prompts in your buyers' own words and run the first sample on each assistant that matters to you. Fix anything that blocks a snippet.

Weeks 3 to 6: the pages

Rewrite the priority pages: the answer first under each question heading, the follow-ups covered, something first-hand added, structured data matched to the text. Ship them in one batch on a recorded date, so the before and after windows are clean.

Weeks 7 to 10: the sources around you

Work through the sources the sampled answers cite on your topics. Update Business Profile and, for retailers, the Merchant Center feed. Make the video or comparison page the answers keep borrowing from someone else, and take part, under your own name, in the communities your buyers use.

Weeks 11 to 13: measure and decide

Compare the changed pages' AI impressions and Web clicks with the control pages over matching windows. Rerun the prompt sample and report the rate with its interval. Keep what moved the changed pages more than the control and drop the rest. Then take the method to the next 20 pages.

AIO for ecommerce websites

For a shop, the product feed matters more than the blog. Google names Merchant Center among the inputs that help products appear in AI responses, and its product structured data documentation says that providing both structured data on web pages and a Merchant Center feed "maximizes your eligibility to experiences." Use merchant listing markup on every page where people can buy, including shipping costs and the return policy.

Merchant Center now offers six optional conversational attributes aimed at AI surfaces such as AI Mode: question and answer pairs, links to documents such as manuals, related products such as accessories and substitutes, an item group title, variant options and a popularity rank. They do not affect product approval, Google says, and details already in the description or product highlights need not be repeated. In its September update for retailers, Google said lululemon's conversational attributes were used in AI Mode recommendations 50% of the time in testing, one example reported by the vendor.

Reviews reach Google's shopping surfaces through the Product Ratings program, which needs at least 50 reviews across your products and depends on accurate identifiers such as GTINs, MPNs and brand names. Fixing those fields is part of the job.

Retailers also get a first-party share of voice against their competitors. Merchant Center's AI performance insights, generally available since September for English-language queries in Australia, Canada, India, New Zealand and the US, compare your share of AI impressions with your competitor set for shopping queries on AI Mode and AI Overviews, one product category at a time. It covers organic traffic only, and Merchant Center sets the competitors, not you.

Checkout is moving into the answer as well. Google's Universal Commerce Protocol puts a Buy button on eligible listings in AI Mode and Gemini, with the merchant remaining the seller of record. It is open to select merchants for now.

AIO tools: what Semrush and the others actually measure

Start with the free first-party reports: Search Console, Bing Webmaster Tools and, for retailers, Merchant Center. They are the only sources that count what the platforms showed. Paid tools add the sampled side.

Semrush spreads its AIO features across several products, according to its knowledge base. Position Tracking records AI Overviews for tracked keywords on any plan, Organic Research shows which of a domain's keywords trigger an AI Overview and whether the domain is cited in it, and Site Audit flags blocked AI crawlers. The AI Visibility Toolkit, from $99 a month, uses generated prompts to report a brand's share of voice and sentiment across Google AI Overviews, AI Mode, ChatGPT, Perplexity and Gemini, updated weekly. Semrush Enterprise AIO, priced on request, adds unlimited prompt tracking for large organizations. The keyword-level AI Overview data comes from Semrush's own checks of results pages; the visibility and share figures are samples.

Ahrefs' Brand Radar works two ways: a prebuilt index of more than 495 million prompts modelled on real searches, which suits comparisons across a market, and custom prompts you write for your own buyers. Its product page lists AI Overviews, AI Mode, ChatGPT, Gemini, Perplexity and Copilot, plus Claude on the index pricing. Custom prompts start at $50 a month and the index at $199.

Before paying for any tracker, ask how many runs sit behind each figure, whether it shows a margin of error, whether you can read the actual answers, and whether its "position" in AI means anything more than the order of a list that changes on every run.

AIO Copilot's AI search view does the sampling for agencies and in-house teams. It reports how often AI assistants name a client against its rivals, with the margin of error shown, and keeps those estimates apart from what Search Console and Bing Webmaster Tools count.

Counted or guessed

The AIO market sells two products under one name: changes to a site, and numbers about what those changes did. The first is mostly the SEO Google has always asked for, rearranged around questions and the sub-questions behind them. The second is where the care belongs.

Search Console now counts your AI impressions and Bing counts your citations, and a retailer's Merchant Center adds a share of voice. Everything else is a sample, and a sample without a denominator and an interval is an anecdote with a chart. Google has said more metrics are coming to its report. When they arrive, the first question to ask of any AIO report will be the same as today: was this counted, or guessed?

Frequently Asked Questions