Answer engine optimization: what AEO is and what can be checked

Four of the first results for the query answer engine optimization opened cleanly on 5 September 2026, and all four define the term the same way. None of them says when the work is not worth doing. None quotes the Google page that states what is actually required to appear in an AI Overview, and the figures they cite arrive without the samples they were measured on. The definition was never the hard part.

What is answer engine optimization?

Answer engine optimization (AEO) is the work of shaping a page so a search or answer engine uses one of its passages as the direct answer. On a new site it wins long-tail questions; on a mid-authority site it is where most of the gain sits; on a consolidated site it is defending snippets already held.

Quick answer

• AEO is four surfaces, not one. Featured snippets, related questions, AI Overviews and spoken answers reward similar writing and pay very differently.
• The eligibility test is one sentence long. Google documents that a page has to be indexed and eligible to be shown with a snippet, and that nothing further is required.
• Three of the four surfaces arrive unlabelled. Clicks count when there is a link to follow, which leaves the spoken answer producing nothing at all.
• Winning can cost the visit. On a query whose answer fits in forty words, the extraction is the transaction.

Why every definition agrees and no two numbers do

The agreement is real and it is also the least useful part.

All four pages describe AEO as structuring content so an answer engine selects and cites it, and then they diverge on everything measurable. One cites 55 % of Google searches carrying an AI Overview, attributed to WordStream, alongside 527 % year-on-year growth in AI-referred sessions from Position Digital and two billion daily ChatGPT queries from Exposure Ninja. Another gives a 14.2 % conversion rate for AI-referred visitors, from Opollo. A third reports that 99 % of URLs shown in AI Mode appear in the top 20 organic results, from Bounteous. None of those figures arrives with the sample it was measured on, and none of them cites a study anyone can open.

Two of the four rest part of their case on a Gartner press release of 19 February 2024, whose headline predicted that traditional search engine volume would drop 25 % by 2026. That target year is the current one. Neither page says what happened.

The model those pages install in a buyer is that AEO is a new discipline with published requirements somewhere, and that compliance is something to purchase. There is a published requirement. It says almost the opposite of what the pitch implies, and it fits in one sentence.

Four answer surfaces, and only one you can count

Grouping them is what makes the advice that follows fit none of them.

A featured snippet lifts a literal passage out of a single page and links back to it. A related questions panel does the same thing in miniature, under a question the reader did not type. An AI Overview composes a summary from several documents and cites some of them. A voice assistant reads one sentence aloud and stops.

Google's Search Console documentation, checked on 5 September 2026, sets out how the first three count. For a featured snippet, "clicking the link to the external page in the featured snippet counts as a click", with standard impression and position rules. For an AI Overview, "clicking a link to an external page in the AI Overview counts as a click", and the position rule is specific: "an AI Overview occupies a single position in search results, and all links in the AI Overview are assigned that same position". AI Mode follows the same methodology as a results page.

Read that carefully and the measurement problem inverts. The common complaint is that AI Overviews cannot be measured. What the documentation describes is the opposite failure: the clicks are already in the report, counted under the same rules as everything else, and the page says nothing about telling them apart. The data is not missing. It is unlabelled.

The spoken answer is the one with no argument to have. A click is defined as following a link, and a read-aloud sentence has no link in it, so nothing enters the report by any route. Where a business decides to work that surface, the return has to be worth having without a number attached, and saying so out loud is more honest than a dashboard that pretends otherwise. Configuring the property that holds the rest of the evidence is an afternoon of work, described in how to set up Search Console.

The three numbers this market runs on

Each of the three is real, and each is smaller than the way it gets quoted.

The click-through figure has two published values and both of them are Ahrefs. In April 2025 Ahrefs reported that the presence of an AI Overview reduces click-through for position 1 by about 34.5 %, from 300,000 keywords split evenly between queries with an AI Overview and informational queries without one, comparing March 2024 against March 2025 on aggregated Search Console data. On 4 February 2026 the same design was re-run over a longer window, December 2023 against December 2025, and reported that an AI Overview now correlates with a 58 % lower average click-through rate for the top-ranking page. Neither figure is a loss measured on a fixed cohort of pages, and the second says as much in its own wording. A deck quoting either one without its date is presenting a moving measurement as a constant.

The Seer figures come as a pair, and the pair is the finding. Seer Interactive, over 3,119 informational queries across 42 organizations and 25.1 million organic impressions between June 2024 and September 2025, measured organic click-through where an AI Overview appears falling from 1.76 % to 0.61 %, which Seer reports as a 61 % decline. Where no AI Overview appears it also fell, from 2.72 % to 1.62 %, a 41 % decline. Click-through is dropping in both conditions, so only the gap between the two declines belongs to the summary. Citing the first number alone inflates the effect by roughly the size of the second.

The 40 % from the academic paper belongs to specific methods, not to a tactic. The GEO study by Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande states that its methods boost visibility by up to 40 % in generative engine responses, and that headline covers the best-performing methods on one metric rather than any single edit anyone can sell. Where those gains land by ranking position is the part that decides whether the work pays, and it is worked through in what generative engine optimization is.

What Google documents as required

One page settles the argument about whether this is something to buy.

Google's documentation on AI features in Search, checked on 5 September 2026, states that "there are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary". It adds that no new machine readable files, AI text files or markup are needed, and that there is no special schema.org structured data to add. Eligibility is defined plainly: a page must be indexed and eligible to be shown in Google Search with a snippet.

The featured snippet page is blunter still. Asked how to mark up a page as a featured snippet, the documentation answers that you cannot: Google systems determine whether a page would make a good featured snippet and elevate it. There is no submission, no request and no markup that grants one.

What the same pages do give is the list of controls that keep a page out, and they are the ones a template inherits by accident: nosnippet, data-nosnippet, max-snippet and noindex. A short max-snippet value is the silent version of the problem, because the page keeps ranking and simply stops appearing in summaries. The documentation also notes that a low max-snippet setting does not guarantee that featured snippets stop. The mechanical checks that catch an inherited directive before it costs a quarter are collected in the basic SEO checklist.

Where the work actually is

Everything above is definition and evidence. What follows is the sequence, and every step in it has a page of its own.

Being in the index before anything else

Every answer surface selects from material a crawler has already stored, which makes indexing the gate in front of the entire discipline rather than a step inside it. A page that Google has not indexed is not a candidate for a snippet, a summary or a spoken answer, and the diagnosis order when that is the situation is in why a page is not indexed.

Writing a block that survives being lifted out

The competing unit is not the page. It is the passage that can be cut out and still say something, which is a property most business writing does not have. The pair below is the same claim written twice.

Before: Which of those two it is depends on the query, and once that is settled the rest follows. Most teams get this backwards, which is why their results look flat. After: A query whose full answer fits in forty words returns the reader to the results page, so winning the box trades the visit for the mention. A query that needs a procedure or a comparison keeps the visit, because the box cannot carry the whole answer.

Two things changed. The opening no longer depends on a paragraph above it, and the assertion became checkable by someone who disagrees. What "most teams get this backwards" contributes to an extraction is nothing, because there is no claim inside it to quote. How that writing layer sits on top of ordinary editorial work is covered in SEO for AI.

Choosing which queries to hand over

This is the decision the surrounding literature skips, and it is per query rather than per site. A query that resolves in one sentence and leaves nothing behind is a donation: the reader is served, the need is gone, and the brand attached to the answer is remembered by nobody. A query whose answer is a procedure, a comparison with conditions, or a judgment that depends on the reader's situation cannot fit in the box, so the summary works as an advertisement for the page underneath it. Sorting real demand into those two piles before writing anything is what free keyword research is for.

Counting what can be counted, and saying so about the rest

The reporting that survives contact with this work has two columns and an admission. Clicks and impressions come from the property, mixed together as the documentation describes. Presence in generated answers comes from asking the systems a fixed set of customer questions repeatedly over weeks and recording who gets named, which is slow, imprecise and still the only method that matches what happens. The admission is the third surface, where nothing is countable. When the motive for all of this is traffic that already fell, the diagnosis belongs before the tactic, and it is in why organic traffic drops.

Where AEO stops and generative work starts

The split that holds in practice is about what the engine does with the text. An answer box takes a literal passage out of one page, so what wins is a block that survives the cut. A generative system writes its own sentences from several documents, so what wins is having something worth borrowing: a number, the name of whoever measured it, and a plain statement of what it means. Where each of those stops being ordinary ranking work is set out in answer engine optimization vs SEO.

ChatGPT sits one layer past that, because a model writing an answer is not extracting a passage but deciding which sources are worth naming at all. The crawler files that determine whether a site is even available to be chosen are a separate question from anything on this page, and the order for checking them is in how to get cited by ChatGPT.

When answer engine optimization is not worth the work

When the page is nowhere near the first results page. Nothing gets elevated out of a position nobody reaches, so the prior job is ranking rather than formatting.

When the demand is local. A business living on searches about who is nearby and open is buying visibility on a surface its customers barely use, and how summaries behave in that specific context is in AI Overviews in local search.

When the content is the product. On a subscription site, giving away the answer gives away the inventory.

When nothing on the site is quotable. If every claim restates someone else's measurement, the prior project is producing a number of your own, which no competitor can copy and no formatting pass can substitute for.

What to expect from each starting point

A snippet is not property. It gets re-evaluated continuously and changes owner often, which makes this maintenance rather than a project with an end date.

The bilingual case deserves its own row and comes with a reservation attached. In two pairs of client sites I operate, same topic in each language, the Spanish half turns impressions into clicks roughly three times better at comparable positions. Why is not separable from those data: English results pages carry more summaries, more ads and more rich results, the query mix differs, and average position moves on its own as that mix changes. The comparison and its four caveats are in Spanish versus English SEO.

Bought rather than built, the work is editorial hours inside a search engagement, since no separate market rate exists for it as a line item. A proposal can still be judged, and the things worth judging it on are which pages get rewritten, which queries get tracked, and who reads the report that comes back. Those three questions separate a vendor selling work from one selling the acronym, and they generalize, which is the subject of how to choose an SEO agency.

The engagement it lives inside is the thing that carries a price, and the spread between the cheapest and the dearest quote for it says more about scope than about quality. Which bands are real, where the centre of the market actually sits, and which variables move a quote inside them, is in SEO cost for a small business.

Underneath both sits the question this article has assumed and never argued: what search work is at all, what it can be expected to move, and how slowly. That is the floor under everything above, and it is in what SEO is.

Mistakes that repeat

• Treating four surfaces as one. They reward similar writing and settle in different currencies, and one of them settles in none.
• Quoting a click-through figure without its date. The same study by the same firm reported 34.5 % in April 2025 and 58 % in February 2026.
• Citing the decline with summaries and not the decline without them. Click-through fell in both conditions, so only the gap is attributable.
• Buying compliance with a specification that does not exist. Google documents no additional requirements and no special markup.
• Pursuing every snippet the site could win. On a one-sentence query the win is a donation with the invoice attached.
• Putting the mention and the session in the same row of a report. One of the two cannot be counted at all, and averaging them hides which is which.

Data and transparency

That no additional requirements or special optimizations are needed to appear in AI Overviews or AI Mode, that eligibility is being indexed and eligible to be shown with a snippet, that no new machine readable files, AI text files or special schema.org structured data are required, and that nosnippet, data-nosnippet, max-snippet and noindex are the controls that limit what is shown, comes from Google's documentation on AI features in Search, opened on 5 September 2026. That a page cannot be marked up to become a featured snippet, that Google systems decide and elevate it, and that a low max-snippet value does not guarantee featured snippets stop, comes from Google's featured snippets documentation, opened the same day. The click, impression and position rules quoted for featured snippets, AI Overviews and AI Mode come from Google's Search Console documentation on the performance report, opened the same day; that page describes how each one counts and gives no method for separating them, which is the basis for calling the data unlabelled rather than missing.

The click-through figures come from two Ahrefs studies of the same design, 300,000 keywords split between queries with an AI Overview and informational queries without one, using aggregated Search Console data: about 34.5 % for position 1 comparing March 2024 with March 2025, published 17 April 2025, and a 58 % lower average click-through rate for the top-ranking page comparing December 2023 with December 2025, published 4 February 2026. Both were opened on 5 September

1. The paired decline figures come from Seer Interactive's study published 4 November 2025, covering 3,119 informational queries across 42 organizations and 25.1 million organic impressions from June 2024 to September 2025: 1.76 % to 0.61 % with an AI Overview, reported by Seer as a 61 % decline, and 2.72 % to 1.62 % without one, a 41 % decline. The visibility figure comes from the GEO paper by Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, arXiv 2311.09735, presented at KDD 2024, and belongs to its best-performing methods rather than to any single tactic.

That four of the first results define AEO the same way, that none of them states when the work is not worth doing, that none quotes Google's eligibility documentation, and that the figures they carry name a publisher without a sample, is my own reading of that results page on 5 September

1. The publishers named in those citations are WordStream, Position Digital, Exposure Ninja, Opollo and Bounteous, reproduced as they appear rather than as verified measurements. The bilingual ratio of roughly three times the clicks at comparable positions is client work, rounded and unnamed, measured across two pairs of sister sites in Search Console; it travels with the reservation stated where it appears, because the candidate explanations cannot be separated with those data. The Gartner forecast of a 25 % drop in traditional search engine volume by 2026 is named from the title of its press release of 19 February 2024 and is not linked here, because the release returned an error to a plain request and an unopened source is not a citation. The search demand for this query, 5,000 monthly searches in the United States with a keyword difficulty of 42 and a cost per click near 6 USD, with an AI Overview present on the results page, was measured in Ahrefs on 5 September 2026 for this blog's keyword research. The operational judgment throughout comes from work across a portfolio recording more than 300 million impressions a year in Search Console. Verified as of September 2026.

Primary sources, opened on 5 September 2026: Google's documentation on AI features in Search; Google's featured snippets documentation; the Search Console performance report reference; the Ahrefs click-through study and its February 2026 update; the Seer Interactive study; the GEO paper (arXiv 2311.09735).

What this changes

The acronym promises a specification and delivers a trade.

There is no requirements document to comply with, no markup that grants an answer and no form to submit, which removes the purchase most of this market is organized around. What is left is a decision nobody outside a business can take for it: which of its own questions it will answer for free, on a results page, in front of a reader who never arrives.

The awkward part is that the decision has to come before the evidence. The clicks this work wins are already sitting in the report, unlabelled and indistinguishable from the ones ordinary rankings bring, so no dashboard is going to hand back a verdict on it later. An honest plan names the queries it intends to give away before any rewriting starts, and treats whatever the report shows afterwards as confirmation rather than proof. Choosing first is not a shortcut around the measurement. It is the part the measurement cannot do.

Frequently asked questions