AI Search Visibility

Start with what the company that runs the feature actually documents

Before evaluating any proposal with GEO or AEO in the title, read the page published by the party that decides. Google's documentation on AI features states that "there are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary."

On eligibility: a page "must be indexed and eligible to be shown in Google Search with a snippet," and "there are no additional technical requirements." On the products being sold hardest right now it is blunter still: "You don't need to create new machine readable files, AI text files, or markup to appear in these features. There's also no special schema.org structured data that you need to add."

That refutes two product categories currently being invoiced, and it is not my opinion about llms.txt and AI-specific markup. It is the stated position of the system you are trying to appear in.

Note what kind of levers the documented controls are. nosnippet, data-nosnippet, max-snippet and noindex restrict or remove snippet eligibility; Google-Extended separately limits AI training and grounding in some of Google's other systems. Every documented control is a way to opt out; there is no documented way to opt in. Google's page on AI features in Search is the source for all of the above.

What became measurable on 3 June 2026, and what did not

On 3 June 2026 Search Console gained a separate view for visibility from generative AI features, reporting impressions, which pages surfaced, countries, devices and dates down to the hour. Appearance is now observable in your own data rather than inferred from someone's panel.

Three limits shipped with it, and they decide what can honestly be reported.

• There is no click data. Without clicks there is no meaningful click-through rate for an AI appearance, and no way to price one.
• AI Overviews and AI Mode are combined into a single category within Search, with a separate report for Discover. They cannot be told apart.
• The rollout was partial — a subset of sites received access first, with no stated timeline for the rest.

So the accurate sentence is: appearance is measurable, the value of that appearance is not. Any provider reporting AI-driven traffic, leads or revenue is not reading it out of Google's data, which does not contain it. They are modeling it. Modeling is legitimate; presenting a model as a measurement is not, and the question separating the two is where does this number come from.

The evidence behind the field, graded honestly

Two pieces of published research carry weight here, and neither says what the pitch decks say.

The founding academic work introduced a benchmark of user queries across multiple domains and reported that its methods "can boost visibility by up to 40% in generative engine responses." Four qualifiers travel with it and are usually stripped off in the retelling. The authors state that efficacy "varies across domains." The benchmark is not Google's production system. The paper was submitted in November 2023, before AI Overviews rolled out in the US. And the per-technique ranking that circulates in summaries is not something I can confirm from the published abstract, so I do not repeat it. The paper is public; read what it claims rather than what it is quoted as claiming.

The largest correlational study examined 75,000 brands across ChatGPT, Google AI Mode and AI Overviews. The strongest associations were YouTube mentions (Spearman correlation around 0.74 for AI Overviews) and branded web mentions. Domain Rating sat far lower, roughly 0.27 to 0.33, and backlinks were weak. The authors attached their own warning, and the first three words of it are the load-bearing ones: "correlation isn't causation." They add that a pattern between a search metric and AI mentions does not mean improving that metric will boost AI visibility. Note the confound in plain sight: each of the top correlates is a proxy for brand size. The study and its caveat are published in full.

The grade, then. One benchmark on non-Google systems with domain-varying results, one correlational study whose authors disclaim causation, and platform documentation stating that no special optimization exists. No technique has been demonstrated on Google's production AI features, replicated, and shown separable from ordinary SEO and brand scale. If one appears, I will say so.

What can be influenced, in descending order of evidence

The entire defensible list, with the reason each item is on it.

• Being indexed and snippet-eligible. Documented as the prerequisite, in those words. Nothing else matters if a page fails this.
• Not blocking yourself. A site carrying an inherited nosnippet, an over-tight max-snippet, or data-nosnippet wrapped around the paragraph that answers the question, is opted out by its own markup. More common than it sounds, and the cheapest thing to check.
• Conventional quality and topical coverage. Documented again, in Google's words: "the best practices for SEO remain relevant for AI features in Google Search." The work that earns a snippet is the work that earns a citation, because it is the same eligibility.
• Answer structure. Writing so that a self-contained answer exists in the first paragraph, and in a question-and-answer block, is defensible on grounds that predate all of this: it is how a passage becomes quotable. It is editorial discipline rather than an AI technique, and it holds up even if no machine ever quotes it.
• Brand presence across the wider web, including video. Correlational only, causation unproven, confounded by size. Worth doing for reasons unrelated to AI surfaces; not worth buying on the strength of a correlation coefficient.
• Deciding on Google-Extended deliberately rather than by accident, since most sites have never made that call at all.

What cannot be promised, and what a proposal promising it is telling you

Nobody can commit to a citation in an AI Overview, a position within one, or a click-through rate from AI features — the last because the click data does not exist to report against. Nobody can commit to a result from a named GEO technique on Google's production systems, because none has been shown to work there and replicated. And nothing is achieved by an llms.txt file or AI-specific schema, on the documented record.

Which makes this unusually easy to evaluate as a buyer. A proposal promising AI Overview placement is either unaware of the documentation or aware of it, and both readings are disqualifying. One selling AI-specific markup is selling what the platform calls unnecessary. A dashboard reporting AI traffic is reporting a model, and the question is what it is modeled from.

The honest version of this service is smaller and less exciting than the version being sold. It is also the only version that survives the client asking, six months in, to see the source.

Prevalence is volatile, and the headline number is not yours

The narrative says AI Overviews are everywhere and always growing. The largest prevalence study, covering more than ten million keywords through 2025, does not support that: trigger rates ran 6.49% in January 2025, peaked at 24.61% in July, and fell to 15.69% by November. Tracking keywords before and after an AI Overview appeared, the same study found zero-click rate decreased, from 33.75% to 31.53%.

Set that beside the much-quoted finding that 68.01% of US Google searches ended without a click between January and April 2026 — a figure that must always carry both caveats, because the panel behind it excludes the Google mobile search app, covering browser searches only, and because the source states that its multi-year comparisons draw on different panel providers and so "aren't the same users or devices." The two do not contradict each other; they measure different things.

Prevalence is category-specific, and it moves. On the keyword set researched for this site, 870 of 1,264 queries returned an AI Overview — 69% — which tells you about this subject area and nothing about yours. The first deliverable is a measurement of your own query set, not a slide carrying someone else's percentage.

What the work consists of, and what lands on your desk

Five steps, and the first two are the ones that get skipped.

1. Establish eligibility. Verify that the pages that matter are indexed, snippet-eligible and not blocked by inherited directives. Unglamorous, and where the actual defects turn up.
2. Baseline the query set. Measure how often AI features appear on the queries that matter to you, with a date attached.
3. Connect the reporting. Where the generative AI report is available, baseline the impressions and the URLs that surface, and state in writing that clicks are not.
4. Fix what is defensible. Eligibility problems, coverage gaps, answer structure — each justified by evidence predating the AI framing.
5. Write down what was ruled out. The list of things not done, with the reason and the source, is what gets handed to the next vendor who calls.

You receive a dated baseline, an evidence-graded work list naming the strength of the case for each item, the ruled-out list, and a statement of which questions your data cannot answer. This is advisory work; the recommendations are implemented by your team or your existing providers.

It is engaged as a one-off review or inside a standing seat. What moves the effort is the number of query classes to baseline, the number of markets, whether the generative AI report has reached your property, and how much eligibility work is already done — not the size of the business.

When you do not need this, and how you would know it worked

You do not need this if the site is not reliably indexed — fix that first, because every AI feature depends on it. Nor do you need it as a separate purchase if a competent SEO program is already running, since most of what can be influenced is already in scope. The useful version there is a short review confirming eligibility and settling which pitches to decline.

Knowing whether it worked is where this field goes wrong. Rising AI feature impressions in Search Console are evidence of appearance and nothing more. It is not traffic and it is not revenue, and an honest report says so on the same line. Eligibility defects fixed and verified is a checkable outcome. So is the count of query classes where your pages surface, against a dated baseline. So, for some buyers, is the money not spent on things the documentation says are unnecessary.

What you will not get is a rising line labeled AI visibility with no stated source. Until Google reports clicks from these features, that line comes from a model, and you are entitled to ask which one.

Frequently Asked Questions

Published 2026-08-11