What changed in 2026 is the controls around AI search, not the ranking machinery. Search Console gained an opt-out for generative AI features, a dedicated performance report appeared, and FAQ rich results were retired for good. Meanwhile Google stated plainly that no new AI-specific file or markup is needed. Here is each one.

Every claim below comes from Google’s own documentation, verified on 8 September 2026. SEO guidance rots — open the primary source again before you implement anything.

1. AI search now has three display requirements

AI Overviews and AI Mode used to have two requirements: the page must be indexable and snippet-eligible. A third has been added — the site must not be excluded from generative AI features in Search Console.

The default is “include”, so nothing breaks if you do nothing. But now that it is a stated requirement, it is worth confirming that an exclusion has not been switched on by accident.

2. Search Console gained a “Search generative AI” setting

Rolled out from 3 June 2026, under Settings → Search generative AI. Three options:

Include your site’s links and content in Search generative AI features (default)
Exclude your site from Search generative AI features
Inherit from the parent property

This matters because it is the only way to opt out of AI features without damaging your normal search snippets. The previous option was nosnippet, which strips the snippet from ordinary search results too and costs you CTR. “Don’t feed the AI” and “don’t show me in search” are finally separate instructions.

Three caveats.

The rollout was staged, but it is finished. Google started with a subset of UK site owners and expanded after testing; the control reached every property worldwide on 31 August 2026. If you still cannot see it, check your permission level (you need owner access) and the property type before assuming anything else.
It is not a ranking signal. Google states the control “isn’t used as a ranking or inclusion signal affecting other parts of Search.” Excluding does not touch normal rankings or Discover.
It is not instant. Exclusion typically takes a few days to propagate. Do not treat it as an emergency removal tool.

3. There is a generative AI performance report

Also from 3 June 2026, the Search generative AI performance report covers AI Overviews and AI Mode, broken down by page, country, device and date. It went worldwide on 31 August 2026, alongside the control. Discover’s generative features live in a separate report — mixing the two up is the easiest way to misread your numbers. Google says it expects to expand the list of surfaces over time.

Two constraints define how you can read it.

No backfill. Data starts shortly before launch, so year-over-year comparison is impossible.
Impressions only, no clicks. For contribution you still depend on the “Web” search type.

The most useful thing to do with it is hunt for informational queries where impressions hold steady while clicks fall. That is the signature of AI Overviews absorbing the answer, and for that cluster of queries the faster move is to optimise for citation rather than for the click.

4. FAQ rich results were fully retired

Retired on 7 May 2026. They had already been limited to government and health sites since 2023; now they are gone entirely. HowTo rich results were retired back in 2023.

The markup is still valid schema.org and carries no penalty if you leave it in place. But no rich result comes out of it, so there is no longer a reason to spend effort adding FAQPage. When you audit an existing site’s structured data, file this under “leave it alone” rather than under “remove it”.

5. Google said AI-specific optimisation is not a thing

In the generative AI optimisation guide published on 15 May 2026, Google wrote that you do not need to create new machine-readable files, AI text files, markup or Markdown to appear in Google Search including its generative AI features — because Google Search does not use them.

That rules out a lot.

Commonly recommended	How Google treats it
Publish an `llms.txt`	Not used. Expect no effect on Google
Chunk content so AI can parse it	Unnecessary
Add AI-specific schema	Structured data is not required for generative AI search
Rewrite in an “AI-friendly” voice	Unnecessary

Generative AI features are grounded in Search’s core ranking and quality systems. There is no separate ranking for AI. Pages that rank well and answer the question clearly get cited. The work is the same work, and there is no shortcut around it.

The accurate framing for llms.txt is: optional, cheap, and thin on evidence. Publish it for other AI vendors if you like, but never sell it as a Google tactic.

6. Treat AI crawlers as two questions, not one

This is where robots.txt goes wrong most often. There are two separate decisions.

May your content train future models? — GPTBot, ClaudeBot, Google-Extended, Applebot-Extended, CCBot. Blocking these costs you nothing in current visibility.
May AI search products cite and link you? — OAI-SearchBot, Claude-SearchBot, PerplexityBot. Blocking these removes you from AI answers and the referral traffic that comes with them.

The usual choice for publishers is to refuse the first and keep the second.

# Refuse training, keep citation in AI search
User-agent: GPTBot
Disallow: /

User-agent: ClaudeBot
Disallow: /

User-agent: Google-Extended
Disallow: /

User-agent: Applebot-Extended
Disallow: /

User-agent: CCBot
Disallow: /

Google-Extended deserves its own note. It is not a crawler but a product token, and the only thing it controls is whether your content may be used for Gemini training and grounding. Blocking it does not remove you from Google Search, does not remove you from AI Overviews or AI Mode, and is not a ranking signal. “Block Google-Extended so the AI stops using my content” is the wrong lever if what you meant was AI Overviews.

Also remember robots.txt is a convention, not access control. User-triggered fetchers such as Perplexity-User are documented as ignoring it.

What did not change

The parts that stayed still are worth checking precisely because trends draw attention away from them.

Core ranking. Generative AI features sit on top of Search’s ranking and quality systems. There is no parallel channel to optimise.
E-E-A-T. A page with a visible author, date and sources is easier for an AI answer to cite. This was not invented for AI search; it simply carries over intact.
Core Web Vitals. LCP under 2.5s, INP under 200ms, CLS under 0.1. INP replaced FID in 2024 and is still current — nothing changed in 2026.

Writing to be cited is writing well

AI answers favour pages that are easy to extract from and easy to attribute. The levers are editorial quality, nothing more exotic.

Put the main answer inside the first 100 words
Give each H2/H3 exactly one question, so the passage stands on its own
Write key facts as extractable sentences: number, unit, date
Include something no other page has — original data, measurements, cases, first-hand experience
Make the author, date and sources verifiable on the page

The last two do the heavy lifting, and the reason is mechanical: a primary source cannot be paraphrased away, so an AI summary has to cite it. Generic content dissolves into the summary and disappears.

Summary

2026 added settings and measurement, not tactics.

Do: check the Search generative AI setting on your property. Use the generative AI report to find queries with flat impressions and falling clicks. Redesign robots.txt around the training-versus-citation split.

Skip: publishing llms.txt as a Google tactic. Adding new FAQPage markup. Building AI-specific schema or an AI-flavoured writing voice.