2026 Search Strategy: Google and AI Assistants | Adviora

Building a 2026 Search Strategy That Covers Google and AI Assistants

Most search strategies still have one destination and one scoreboard. They aim at Google's results page, and they are judged on position. That was a reasonable simplification for a long time. It is now a strategy with a hole in it, because a growing share of the questions your buyers ask are answered somewhere your rank tracker cannot see.

The instinct is to bolt on a second programme called GEO and run it alongside SEO. That is expensive, duplicative and usually unnecessary. A better model treats one well-built content asset as the input to four different surfaces, each with its own eligibility rules, its own selection mechanism and its own measurement.

This session sets out that framework, walks through the crawler-access decision that quietly determines whether assistants can cite you at all, and ends with a quarterly measurement plan you can hand to a team. Adviora's Executive Dashboard is the reporting layer we use to keep the four surfaces visible in one place.

Why a 2026 Search Strategy Cannot Be a Google-Only Strategy

The argument is not that Google is declining. It is that Google is no longer the only place your content has to be eligible, and eligibility elsewhere is not automatic.

• Different crawlers. Assistants build their own indexes with their own user-agents; your Googlebot access grants them nothing.

• Different selection. Google ranks a list. An assistant retrieves a handful of passages and synthesises one answer, then attributes some sources.

• Different queries. Assistant prompts are longer, messier and more conditional than typed queries, so they touch parts of your topic that keyword tools never surfaced.

• Different reporting. Three of the four surfaces have no first-party analytics at all.

A strategy that only optimises for the surface it can measure will drift towards the surface it can measure. The framework below is designed to stop that drift.

The Framework: One Asset, Four Surfaces

Build the asset once, then check it against four eligibility tests. Every page in a priority cluster should be able to answer all four questions with a yes.

• Blue links - is this page indexed, technically sound, internally linked, and the best available answer for its query?

• Answer boxes - does it contain a self-contained passage in the shape the question implies, positioned directly under a heading that asks it?

• AI Overviews and AI Mode - is it snippet-eligible, readable as text rather than locked in images or scripts, and consistent between its structured data and its visible content?

• Third-party assistants - can OAI-SearchBot, Claude-SearchBot and PerplexityBot actually fetch it, and does it contain claims specific enough to be worth citing?

Notice that questions one to three are largely the same work. The fourth is where most organisations have a silent failure, because nobody owns robots.txt at the strategy level.

Surfaces One and Two: Ranked Links and Answer Boxes

These two share an index and diverge only in presentation. Ranking gets you into consideration; extraction is decided separately and is not something you can request.

• You control eligibility. Google's snippet controls - nosnippet, data-nosnippet and max-snippet - restrict extraction, and restricting it also reduces eligibility for AI features.

• You control shape. Definition sentence for “what is”, ordered list for “how to”, comparison for “X vs Y”.

• You do not control selection. Google's documentation is explicit that site owners cannot influence which page becomes a featured snippet.

• You do control the trade-off. Winning an answer box can lower clicks on that query while raising exposure - decide, per cluster, which you want.

That trade-off is the one to bring to a leadership meeting, and it is explored in Why Ranking #1 No Longer Means Being Found.

Surface Three: AI Overviews and AI Mode

This is where the industry has generated the most noise and Google has been the most direct. Google's Search Central documentation on AI features states that there are no additional requirements to appear in AI Overviews or AI Mode, and no other special optimisations necessary. Eligibility means being indexed, being eligible to appear in regular Search with a snippet, and meeting Search's technical requirements.

• Not required: new machine-readable files, AI-specific markup, or special schema.org structured data reserved for AI features.

• Recommended: allow crawling, make content findable through internal linking, deliver a good page experience.

• Recommended: present important content as readable text, and keep structured data consistent with what is visible on the page.

• Implication: your AI Overviews plan is your technical SEO plan, plus better passages.

The honest reading is that eligibility is cheap and selection is not guaranteed. Anyone selling a proprietary file or markup for this surface is selling something Google says it does not use.

Surface Four: Third-Party Assistants and the Crawler Access Decision

This is the single most useful thing in the session. Each AI vendor runs more than one crawler, and they do different jobs. Blocking them as a group is a decision about training that silently destroys visibility.

• OpenAI - GPTBot crawls for training foundation models. OAI-SearchBot powers ChatGPT's search features, and OpenAI states that blocking it means the site will not appear in ChatGPT search answers. ChatGPT-User covers user-initiated fetches, where robots.txt rules may not apply because a human triggered the request.

• Anthropic - ClaudeBot crawls for training, Claude-SearchBot indexes to improve search result quality, and Claude-User handles user-directed retrieval. Anthropic states its bots honour industry-standard robots.txt directives.

• Perplexity - PerplexityBot surfaces andadhu links sites in Perplexity results and is not used for model training. Perplexity-User covers user-initiated visits and generally ignores robots.txt because a real person initiated them.

• Google - Google-Extended controls whether content may be used for training Gemini models and grounding in Gemini Apps. Google states it does not impact a site's inclusion in Google Search and is not used as a ranking signal.

So the decision is not “allow AI or block AI.” It is a grid: training yes or no, per vendor; retrieval yes or no, per vendor. Adviora's GEO & AEO Visibility module includes an AI crawler access audit that reports the grid as it actually stands, and our AI Crawler Access Audit article walks through the robots.txt syntax.

Filling In the Crawler Decision Grid

Run this as a 30-minute exercise with legal, brand and engineering in the room. It produces a defensible policy instead of an accident.

• List every subdomain. Rules must be applied to each host you want covered; a policy on the main domain does not cover a help centre on another subdomain.

• Decide training access per vendor. This is a licensing and brand question, not an SEO one. Blocking training does not reduce Google rankings.

• Decide retrieval access per vendor. Blocking a search crawler removes you from that assistant's cited answers - treat it as switching off a channel.

• Prefer robots.txt over IP blocking. Anthropic notes that blocking by IP means the bot cannot read robots.txt at all.

• Verify rather than assume. Check server logs for each user-agent, and re-check after any CDN, WAF or platform migration.

• Record the decision and the date, so the next team does not silently reverse it.

The commonest finding is a blanket disallow added years ago by someone protecting bandwidth, still blocking the retrieval crawlers that would otherwise cite the brand.

The Passage-Level Writing Discipline

Surfaces two, three and four all select passages rather than pages. That single fact should change how briefs are written. The unit of work is no longer the article; it is the paragraph that could be lifted out of it.

• One idea per passage, with the subject named in full. A passage that begins “it depends on” cannot be quoted.

• Put the answer first and the reasoning second - the inverted pyramid, applied at paragraph scale.

• Attach something specific: a threshold, a count, a named method, a date. Specifics survive summarisation; adjectives are dropped.

• Match the heading to the question, then answer it in the next sentence rather than three paragraphs later.

• Keep claims consistent across pages so retrieval sees one version of your fact, not three.

This discipline is also what makes content easier to reuse in decks, sales enablement and help documentation, which is usually how you sell it internally.

Why a 2026 Search Strategy Cannot Be a Google-Only Strategy

The argument is not that Google is declining. It is that Google is no longer the only place your content has to be eligible, and eligibility elsewhere is not automatic.

• Different crawlers. Assistants build their own indexes with their own user-agents; your Googlebot access grants them nothing.

• Different selection. Google ranks a list. An assistant retrieves a handful of passages and synthesises one answer, then attributes some sources.

• Different queries. Assistant prompts are longer, messier and more conditional than typed queries, so they touch parts of your topic that keyword tools never surfaced.

• Different reporting. Three of the four surfaces have no first-party analytics at all.

A strategy that only optimises for the surface it can measure will drift towards the surface it can measure. The framework below is designed to stop that drift.

The Framework: One Asset, Four Surfaces

Build the asset once, then check it against four eligibility tests. Every page in a priority cluster should be able to answer all four questions with a yes.

1. Blue links - is this page indexed, technically sound, internally linked, and the best available answer for its query?

2. Answer boxes - does it contain a self-contained passage in the shape the question implies, positioned directly under a heading that asks it?

3. AI Overviews and AI Mode - is it snippet-eligible, readable as text rather than locked in images or scripts, and consistent between its structured data and its visible content?

4. Third-party assistants - can OAI-SearchBot, Claude-SearchBot and PerplexityBot actually fetch it, and does it contain claims specific enough to be worth citing?

Notice that questions one to three are largely the same work. The fourth is where most organisations have a silent failure, because nobody 

Surfaces One and Two: Ranked Links and Answer Boxes

These two share an index and diverge only in presentation. Ranking gets you into consideration; extraction is decided separately and is not something you can request.

• You control eligibility. Google's snippet controls - nosnippet, data-nosnippet and max-snippet - restrict extraction, and restricting it also reduces eligibility for AI features.

• You control shape. Definition sentence for “what is”, ordered list for “how to”, comparison for “X vs Y”.

• You do not control selection. Google's documentation is explicit that site owners cannot influence which page becomes a featured snippet.

• You do control the trade-off. Winning an answer box can lower clicks on that query while raising exposure - decide, per cluster, which you want.

That trade-off is the one to bring to a leadership meeting, and it is explored in Why Ranking #1 No Longer Means Being Found.

Surface Three: AI Overviews and AI Mode

This is where the industry has generated the most noise and Google has been the most direct. Google's Search Central documentation on AI features states that there are no additional requirements to appear in AI Overviews or AI Mode, and no other special optimisations necessary. Eligibility means being indexed, being eligible to appear in regular Search with a snippet, and meeting Search's technical requirements.

• Not required: new machine-readable files, AI-specific markup, or special schema.org structured data reserved for AI features.

• Recommended: allow crawling, make content findable through internal linking, deliver a good page experience.

• Recommended: present important content as readable text, and keep structured data consistent with what is visible on the page.

• Implication: your AI Overviews plan is your technical SEO plan, plus better passages.

The honest reading is that eligibility is cheap and selection is not guaranteed. Anyone selling a proprietary file or markup for this surface is selling something Google says it does not use.

Surface Four: Third-Party Assistants and the Crawler Access Decision

This is the single most useful thing in the session. Each AI vendor runs more than one crawler, and they do different jobs. Blocking them as a group is a decision about training that silently destroys visibility.

• OpenAI - GPTBot crawls for training foundation models. OAI-SearchBot powers ChatGPT's search features, and OpenAI states that blocking it means the site will not appear in ChatGPT search answers. ChatGPT-User covers user-initiated fetches, where robots.txt rules may not apply because a human triggered the request.

• Anthropic - ClaudeBot crawls for training, Claude-SearchBot indexes to improve search result quality, and Claude-User handles user-directed retrieval. Anthropic states its bots honour industry-standard robots.txt directives.

• Perplexity - PerplexityBot surfaces and links sites in Perplexity results and is not used for model training. Perplexity-User covers user-initiated visits and generally ignores robots.txt because a real person initiated them.

• Google - Google-Extended controls whether content may be used for training Gemini models and grounding in Gemini Apps. Google states it does not impact a site's inclusion in Google Search and is not used as a ranking signal.

So the decision is not “allow AI or block AI.” It is a grid: training yes or no, per vendor; retrieval yes or no, per vendor. Adviora's GEO & AEO Visibility module includes an AI crawler access audit that reports the grid as it actually stands, and our AI Crawler Access Audit article walks through the robots.txt syntax.

Filling In the Crawler Decision Grid

Run this as a 30-minute exercise with legal, brand and engineering in the room. It produces a defensible policy instead of an accident.

1. List every subdomain. Rules must be applied to each host you want covered; a policy on the main domain does not cover a help centre on another subdomain.

2. Decide training access per vendor. This is a licensing and brand question, not an SEO one. Blocking training does not reduce Google rankings.

3. Decide retrieval access per vendor. Blocking a search crawler removes you from that assistant's cited answers - treat it as switching off a channel.

4. Prefer robots.txt over IP blocking. Anthropic notes that blocking by IP means the bot cannot read robots.txt at all.

5. Verify rather than assume. Check server logs for each user-agent, and re-check after any CDN, WAF or platform migration.

6. Record the decision and the date, so the next team does not silently reverse it.

The commonest finding is a blanket disallow added years ago by someone protecting bandwidth, still blocking the retrieval crawlers that would otherwise cite the brand.

The Passage-Level Writing Discipline

Surfaces two, three and four all select passages rather than pages. That single fact should change how briefs are written. The unit of work is no longer the article; it is the paragraph that could be lifted out of it.

• One idea per passage, with the subject named in full. A passage that begins “it depends on” cannot be quoted.

• Put the answer first and the reasoning second - the inverted pyramid, applied at paragraph scale.

• Attach something specific: a threshold, a count, a named method, a date. Specifics survive summarisation; adjectives are dropped.

• Match the heading to the question, then answer it in the next sentence rather than three paragraphs later.

• Keep claims consistent across pages so retrieval sees one version of your fact, not three.

This discipline is also what makes content easier to reuse in decks, sales enablement and help documentation, which is usually how you sell it internally.

What to Measure Quarterly

Weekly measurement of AI visibility produces noise and arguments. Quarterly measurement, on a fixed method, produces decisions.

• Search performance - clicks, impressions, CTR and average position from Search Console, split into informational and commercial query sets.

• Answer-box presence - whether your priority questions are being resolved on the results page, sampled manually on a fixed query list.

• Citation coverage - the share of a fixed prompt set where your brand is cited, recorded per assistant, alongside a named competitor set.

• Crawler access - a pass/fail control check that the intended grid is still in force after any infrastructure change.

• One qualitative read - how your brand is described when it is cited, since sentiment matters more when the model is speaking on your behalf.

Adviora's Executive Dashboard carries these as separate tiles, and the definitions behind each metric are in our Glossary of Modern Search Metrics. For the wider leadership view, see The CMO's Marketing Intelligence Dashboard.

Best Practices

1. Write the strategy as one asset against four eligibility tests, not as two competing programmes.

2. Make the crawler grid an explicit, dated, signed-off decision with legal and engineering present.

3. Never block a retrieval crawler to protect training data; block the training crawler instead.

4. Rewrite briefs at passage level, with the answer first and a specific fact attached.

5. Keep structured data consistent with visible content on every template.

6. Fix the measurement method before the first measurement, and change it as rarely as possible.

7. Report exposure and demand as separate lines to leadership.

The Future of Multi-Surface Search

Expect more surfaces, not fewer, and expect the access layer to keep formalising. Vendors already publish IP ranges and distinct user-agents; proposals such as llms.txt are circulating without any major engine committing to honour them. The durable investment is not a file or a markup trick - it is a content estate whose claims are specific, consistent and reachable, and a policy that says deliberately who may read it and for what purpose.