Staying in AI answers

Do we want to appear in AI-generated answers? If so, on what terms? Who decides?

The levers available

• From the outset, robots.txt can instruct AI search engines, or some of them, whether to scrape information.
• Since 17 June 2026, Search Console has also allowed sites to opt out of AI Overviews and AI Mode without an organic ranking penalty.

Brands therefore face a choice that concerns not only SEO but the business itself.

The question

Do we want to appear in generated answers, and on what terms? And, before that: who decides?

Why it is a decision, and why now

Since 17 June 2026, Search Console has offered a toggle to opt out of AI Overviews and AI Mode without penalising organic ranking. The decision therefore needs to be made and documented rather than left to omission.

Before that date the topic was analytical: organisations observed a phenomenon. Since then, it is a checkbox with a default value, and the default value is someone else’s choice.

What the literature says

• The cost of presence. CTR down between 15% and 89% depending on the source, 93% zero-click in AI Mode, Reuters at −33% and −38%.
• The value of presence. 0.5% of sessions produce 12.1% of signups: roughly 23 times the average rate.
• The instability. 42% of cited domains get replaced after a change of default model; the presence of AI Overviews on e-commerce queries drops from 29% to 3.2%. The point is not only “stay or leave”: it is how much to depend on a channel that reconfigures itself without notice — so when it is worth staying in AI answers, and when it is worth handing control back to our own site.
• What has become a lever. Preferred Sources, since 27 May 2026, with over 345,000 sources already selected by users. Here the choice is made by the reader: once a source is preferred, ranking matters less than explicit preference, so the lever is not just ranking but the quality of the relationship the reader builds before reaching the result.

The cost of either branch

Staying: the organisation feeds a channel that intercepts its audience before the site, with visibility that does not convert into traffic and metrics that since 3 June 2026 give impressions only. In parallel, it is just as important to weigh what that AI traffic is actually worth: What AI traffic is worth.

Leaving: the organisation also leaves the comparison, and its representation gets built by earned sources — which are 99% of them.

Not deciding: you stay exposed to the vendor’s default, today “everything open”, without distinguishing between indexing, visibility in answers, and use for training.

Criterion for deciding

Separate the decision on access and training from the decision on presence in answers. Weigh together the channel’s composition and conversion, its dependence on an unstable surface, and the quality of the relationship that leads to Preferred Sources — not a traffic time series that is no longer available.

The choice also includes the quality of representation. Only 1% of sources cited by LLMs comes from brand-owned sites: staying visible does not replace work on earned sources, and leaving places almost the whole representation with third parties. When a model attributes false information to a brand, monitoring moves discovery earlier but offers no direct correction; the investment and legal-escalation threshold therefore needs to be declared with the decision.

How it is measured

Badly, and this has to be said. Since 3 June 2026 the Generative AI reports in Search Console give impressions only: the figure that would allow a before-and-after comparison is gone. The window for building a comparable time series has already closed, so the decision has to be made on composition and conversion data, not on a traffic series. In practice we can no longer read this choice as a linear before/after: we have to read it as a decision about the channel’s composition.

Who decides

SEO, GEO and AI Visibility owns the decision because it touches content, SEO/GEO, brand and channel budget at once. The corpus names the AI Visibility Analyst — a discipline that is anything but marginal, yet rarely exists as a role in most organisations; see SEO, GEO and AI Visibility. If that figure is not there, the decision has no groundwork behind it: there is only the checkbox.

An upstream decision: access and training

Staying in or leaving AI answers is not the first decision to make: upstream of it sits the choice of whether to let AI crawlers access your content at all, and whether they can train models on it — two different things, often collapsed into a single robots.txt line.

The levers are finer-grained than they look. robots.txt is still the starting point, but a single blanket rule is no longer enough: Google-Extended separates Gemini’s training from Search indexing, so it is possible to stay visible in Search while opting out of training — two separate permissions, not one. Bot by bot: GPTBot, ClaudeBot, PerplexityBot and OAI-SearchBot each answer to their own directives.

Cloudflare, which serves roughly 20% of the web, changed the default: for new domains AI crawlers are blocked out of the box, and Pay Per Crawl revives the HTTP 402 (Payment Required) status so publishers can choose, request by request, to Allow, Charge or Block. RSL (Really Simple Licensing, launched in September 2025 with Yahoo and Medium, implemented via robots.txt) does the same job in the form of a declared licence.

On llms.txt: it is not a visibility lever, and it is worth saying plainly because the opposite still circulates. Proposed by Jeremy Howard in September 2024, it is not supported by Google — Gary Illyes said so explicitly at Search Central Live (July 2025), and John Mueller compared it to the keywords meta tag, a tool the industry already wrote off long ago. No major LLM provider reads it in production: a 2026 analysis of over 500 million AI-bot visits found requests to /llms.txt statistically negligible, and roughly 40% of published files are plugin-generated defaults rather than purpose-written content. Perplexity, Claude and coding agents do read it, and it costs little to publish for developer documentation — but it is not a ranking factor in AI search, and treating it as one moves effort from where it counts to where it does not.

Blocking training does not mean leaving AI answers — Google-Extended proves that by separating the two permissions — but the choice has to be made with the same awareness as this card: whoever does not decide stays on the vendor’s default, which today is “everything open”. See also item 8 of The regulatory perimeter on access policy for automated agents — the same family of levers, applied to agents that act rather than crawlers that learn.

Where it falls on roles

Staying or leaving is not a configuration: it assigns different work to different functions, including one that does not exist in most organisations.

• SEO, GEO and AI Visibility — Owns the decision; the work shifts from position to share of presence measured on versioned prompt sets across platforms.
• Content & Editorial — Citation becomes a distinct goal from ranking: self-contained paragraphs, informative headings and verifiable, current data.
• Brand & PR — Earned sources are the infrastructure of visibility; if the organisation leaves, they become the only one.
• Marketing & Growth — Defends the channel on composition and conversion, not volume.

Related phenomena

• Representation forms beyond owned channels — Only 1% of cited LLM sources come from brand sites.
• The customer journey shortens — Awareness and consideration happen in chat.
• Intent-based interaction — Visibility is built around intent, not keywords.
• The conversations you cannot observe — The choice cannot be fully validated after the fact.
• The regulatory perimeter — Access and training levers are positions the organisation answers for.