Gemini AI Mode Update: What Changed and How to Stay Visible in AI Answers

From chat experiment to a permanent search surface

AI Mode started as a Labs experiment in March 2025, became generally available in the US around Google I/O in May 2025, and has since been treated as a standing feature rather than a trial. It sits alongside the classic results tabs, accepts long conversational prompts, and keeps context for follow-up questions.

The model layer behind it has moved too. Google announced Gemini 3 on 18 November 2025 and said it was bringing the model into AI Mode in Search from launch, initially for its paid Google AI subscribers before wider availability. The practical effect of newer models is longer, more reasoned answers with more sub-topics covered inside a single response, which means more opportunities for a brand to be mentioned and more opportunities to be left out.

What UK searchers currently see

Google announced AI Mode for UK users in late July 2025, in English, following the earlier US rollout and a wider international expansion. Since then Google has reduced friction to reach it: the tab is prominent on mobile, and mobile users tend to encounter AI-generated answers first because the response fills the screen before any organic listing appears. On desktop, classic results are still doing more of the work for many queries.

Availability shifts frequently, so treat any specific claim about what UK users see as something to verify by testing rather than something to read once. That is a habit worth building for its own sake.

The capability additions worth caring about commercially

• Multi-step research responses. AI Mode will break a broad question into parts and answer them in sequence, citing different sources for different parts.

• Follow-up questions. The second and third prompt in a session are where shortlists form, and they are invisible to conventional keyword research.

• Agentic tasks. Google has been extending AI Mode towards booking and task completion for some users, which shortens the path from question to decision.

• Shopping-style comparisons. Structured comparisons of options, including product and provider attributes pulled from multiple pages.

Why consideration-stage queries feel it most

Navigational searches are largely unaffected. Someone typing your brand name still gets you. The pressure lands on consideration-stage language: "best X provider for Y", "alternatives to Z", "which supplier suits a mid-size UK firm". Those are the queries where a synthesised answer genuinely replaces the click, because the reader wanted a shortlist rather than a page.

These get treated as one thing in most coverage, and that is the root of a lot of wasted work. AI Overviews sit above traditional results for a subset of queries. AI Mode is a separate conversational surface inside Search. The Gemini app is a standalone assistant that may or may not be grounding a given answer in Google Search at all.

Because they ground answers differently, the same brand can be described accurately in one and miscategorised in another. We have seen a security company positioned correctly as a threat intelligence provider in one surface and filed under a neighbouring category in another, off the back of a handful of third-party pages that described it loosely. Nothing on the company's own website was wrong. The sources being cited were.

Which controls apply where

This is the most consistently misunderstood part of the update. The Google-Extended user agent token controls whether your content is used to help ground Gemini app responses and Vertex AI grounding. It is not an AI Mode opt-out. Answers grounded in Google Search, including AI Overviews and AI Mode, are governed by the standard snippet controls: the nosnippet and max-snippet robots directives and the data-nosnippet HTML attribute. Google's own robots meta tag documentation sets out how these behave.

The trade-off is the point. Applying nosnippet to keep content out of AI answers also strips the classic snippets and rich results you almost certainly want to keep, and it does nothing to stop a competitor's page, a directory listing or a review site from being cited instead. Blocking is rarely the right lever. Being clearer than the alternatives usually is.

Testing one surface and assuming the rest match

If you check AI Overviews for ten queries and conclude you are fine, you have measured one surface on one device in one session. Fixes confined to your own website tend to move one surface. Changing the third-party sources being cited tends to move several, because those sources are shared across grounding methods.

Multilingual and multi-market sites

Language rollout is uneven, so a UK English answer and a Brazilian Portuguese answer for the same product can be grounded in different source sets. Multi-market teams should build separate prompt sets per market rather than translating one set and assuming parity. Local directories and regional review sites carry more weight than most global brands expect.

Google uses the term "query fan-out"for the behaviour at the centre of AI Mode: one prompt is decomposed into multiple related background searches, and the answer is assembled from the results of all of them. You never see those queries.

Two consequences follow. First, a page can be cited for a sub-question it has never ranked for as a standalone term. Second, a page can rank first for the visible query and appear nowhere in the answer, because the fan-out queries went somewhere else. Rank tracking measures a unit the system no longer uses on its own.

Passage-level retrieval punishes buried answers

Retrieval happens at passage level. A 3,000-word page that answers the question properly in paragraph 24, hedged across two sentences, loses to a competitor page that answers it plainly under a clear heading. This is why "we already have content on that"is not an answer. The question is whether a specific, quotable passage exists, near a heading that matches the question, in language a model can lift without interpretation.

One screenshot is anecdote

AI Mode answers vary between sessions, devices and personalisation states. Run the same prompt three times and you may get three different citation sets. So treat visibility as a sampled percentage across repeated runs, not a binary present-or-absent check, and only claim a change when the same prompt set moves consistently over time. A screenshot in a board deck proves nothing except that one session existed.

Turning keywords into buyer prompts

The workable conversion is from a shortlist of head keywords into 15 to 100+ prompts phrased the way buyers actually ask. From "penetration testing", you get:

• "best penetration testing providers for UK fintechs"

• "who should I use instead of [competitor]"

• "which pen test provider handles FCA-regulated clients"

• "how much does a CREST-accredited pen test cost in the UK"

Those prompts become the measurement unit. Run them daily, record which brands are named and which sources are cited, and you have something you can trend. This is the basis of the AI visibility plans and pricing we run for clients, and it is the part most teams can start without us.

AI Mode activity is not broken out as a separate channel in Search Console. Google aggregates it into the Web search type in Performance reports, so there is no dimension to filter on and no way to isolate it cleanly.

The pattern to look for instead: impressions holding steady while clicks and click-through rate slide on informational and comparison queries, with branded and transactional queries broadly unchanged. That signature is what AI Overviews traffic loss usually looks like in the data you already have. Teams without intent cohorts in place read it as a seasonal dip and wait for it to recover.

Build intent cohorts before you need them

Segment your queries into four buckets, informational, comparison, branded and transactional, and record a baseline CTR for each. When a change lands, you can see which cohort moved. Without that split, aggregate CTR averages the damage away, because branded traffic holding up masks informational traffic falling.

What you can and cannot attribute

Some assistant-driven visits arrive with identifiable referrers and can be segmented in analytics. Many do not, and volumes are typically small relative to the influence involved, because the assistant answered the question. Be honest about the boundary. A visibility score can show whether your brand is being named, how often, in what terms and alongside which competitors. It cannot prove a revenue figure, and anyone presenting it as attributed pipeline is overreaching.

Start with source attribution, not your own site

Before rewriting anything, collect the sources actually cited in the answers you care about. In most B2B categories the list is dominated by third-party pages: directories, "top 10"listicles, review platforms, trade publications, comparison sites. That list tells you where the answer is coming from. Working only on your own website while ignoring the sources being quoted is the most common expensive mistake in this space.

One owned page rarely flips a category-level answer on its own. Corroboration across several independent sources usually does, because the model is looking for agreement rather than assertion.

Apply the six pillars in order

1. Clarity. Say what you do, for whom, in which country, in plain sentences a model can quote.

2. Consistency. The same category description everywhere: site, LinkedIn, Companies House, directories, press coverage.

3. Trust. Named authors, credentials, accreditations, verifiable client references where permitted.

4. Visibility. Presence in the third-party sources already being cited in your category.

5. Freshness. Dated updates on the pages that answer comparison questions.

6. Technical foundations. Crawlability, structured data, clean headings, no accidental snippet blocking.

The order matters. Authority building on top of an unclear category description just spreads the wrong description faster.

Fixing misrepresentation

The method is unglamorous and it works. Document the wrong description verbatim, with the prompts and dates that produced it. Identify the sources feeding it. Change those signals, on your own site and on the third-party pages you can influence. Then retest the same prompts on the same schedule and compare. That was the shape of the DarkInvader work: a brand AI systems were slotting into the wrong category, corrected through representation and source changes, then re-tested against the original prompt set to confirm the description had genuinely moved rather than assuming it had. Confirmation is the step people skip.

Model updates will keep landing. A routine built on prompts and sources survives them; a routine built on rank reports does not.

• Daily prompt checks across the assistants your buyers actually use: Gemini, ChatGPT, Perplexity, Claude, Copilot and Meta AI. Same prompts, same time, recorded results.

• Weekly category leaderboard. Who is being recommended in your category this week, and which cited sources do they appear in that you do not.

• Monthly source review. New directories, listicles and review pages entering the citation set.

Governance framing that stands up internally

If you need this reporting to survive scrutiny from risk or compliance colleagues, map it to the NIST AI Risk Management Framework functions. Govern: who owns AI representation of the brand. Map: which surfaces and prompts are in scope. Measure: the sampled visibility percentages and citation sources. Manage: the remediation backlog and retest schedule. It converts an interesting marketing chart into a documented process.

What implementation actually involves

Remediation typically splits four ways: technical fixes, content clarity rewrites, brand representation work across third-party sources, and authority or citation building. Sequencing varies by how badly the current answers misdescribe you. We deliver this on a day-rate basis at £750 per day rather than as a single content push, because the work is iterative and the retest cycle is part of it. If you want a smaller starting point, a snapshot audit of how AI systems currently describe you gives you the baseline before you commit to anything, and a growth consultation is the right route if you already know the answers are wrong and need a plan.

The short version of the Gemini AI Mode update: nothing about it makes traditional SEO irrelevant, but it does make prompt-level sampling and source attribution the measurements that explain what is happening. Generative engine optimisation in the UK is still a young discipline, and the teams doing well at it are the ones treating it as measurement discipline rather than a content sprint.

What is Gemini AI Mode in Google Search?

AI Mode is a conversational search surface inside Google, powered by Gemini models, that accepts long questions and returns a synthesised answer with links, then supports follow-up questions in the same session. It began as a Search Labs experiment in March 2025 and is now a standing feature with its own entry point in Search.

Is the Gemini AI Mode update available to UK searchers?

Yes. Google announced AI Mode for UK users in English in late July 2025, following the earlier US rollout, and has continued expanding availability and language support since. Because access and features change often, verify what your own audience sees by testing on both mobile and desktop rather than relying on a published rollout list.

Does AI Mode use the same ranking signals as normal Google results?

It is grounded in Google Search, so the same indexing and quality foundations matter, but the retrieval process differs. Query fan-out means one prompt triggers multiple background searches, and retrieval happens at passage level, so citations often land on pages that do not rank first for the visible query.

Can I stop AI Mode from using my content, and should I?

The Google-Extended token does not control AI Mode; it applies to Gemini app grounding and Vertex AI grounding. Search-grounded answers are governed by nosnippet, max-snippet and data-nosnippet, which also remove the standard snippets and rich results you probably want to keep. Blocking rarely helps, because competitors and third-party pages will simply be cited instead.

Why can't I see AI Mode traffic separately in Google Search Console?

Google aggregates AI Mode data into the Web search type in Performance reports rather than providing a separate dimension. The usual way the effect shows up is stable impressions with falling clicks and CTR on informational and comparison queries, while branded and transactional queries hold steadier.

How do I get my brand cited in Gemini AI Mode answers?

Identify the sources currently cited for your priority prompts, then work on being clearly and consistently described in those sources as well as on your own site. Clear, quotable passages under matching headings help, but corroboration across several independent third-party sources tends to move category-level answers more reliably than any single owned page.

How often should I retest prompts after an AI Mode update?

Daily sampling gives you a trend you can trust, because answers vary between sessions and users. After any model or feature change, run your existing prompt set for at least a week or two before drawing conclusions, and only report a shift when the same prompts move consistently rather than on the strength of one screenshot.