If you’re still debating whether AI search is something to watch rather than act on, the data is making that harder to hold. Right now, 30% of consumers use AI for product research. A year ago, that figure was 12%. And roughly 40% of Google results now include an AI Overview, meaning people are reading AI-generated answers without choosing to.
How many people are actually using AI search?
Add those two things together and the share of search journeys that involve AI in some form is already closer to 60%. Brainlabs projects it will reach 80% within twelve months, based on current growth rates holding. That changes how you think about almost everything in organic search.
Brainlabs survey data splits consumers into three groups. Traditionalists (around 69%) use established search engines exclusively. Augmenters (around 30%) use both Google and AI platforms at different points in a single research journey. Dissenters (fewer than 1%) use only AI platforms. The augmenters are where the growth is — they haven’t abandoned Google, they’ve added AI as a research layer for complex queries.
Why AI SEO is not just a rebrand of traditional SEO
The honest answer: similar enough right now that you don’t need to tear up your roadmap, but different enough that the gap is widening faster than most teams have adjusted for. There isn’t one AI search — there are at least four major platforms with different models, different guidelines, and a different relationship with Google’s index.
That 15% figure for Gemini is the one worth pausing on. Gemini is a Google product, yet it only cites pages from Google’s top 10 results 15% of the time. An SEO strategy built entirely around top-10 rankings does very little for Gemini or ChatGPT. And the overlap keeps shifting — AI Overviews used to have a 76% overlap with Google’s top 10. In 2026, it has halved.
How LLMs actually decide what to cite
Understanding why an obscure page ends up cited in a Gemini answer while a major brand’s content is ignored requires understanding how LLMs generate responses. It’s genuinely different from how Google ranks pages.
Grounding: When Gemini receives a prompt, it asks itself whether it needs external data or can rely on training data alone. For most product research queries, it needs external data.
Query fan-out: The LLM generates dozens of related micro-questions from the original prompt. For complex prompts there can be tens of these running simultaneously.
Deep index search: For each micro-query, the LLM searches Google’s index across the top 100 results and beyond. Not just the top 10 — this is the first major divergence from traditional SEO.
Content selection: The LLM looks for direct answer blocks, headings matching the micro-question, hard statistical data, and freshness signals like a recent “last updated” date.
Comparison and validation: Sources with low consensus compared to higher-authority references get filtered out. A page that answers a specific micro-question very directly can outcompete higher-authority pages that bury their answer in editorial prose.
What LLMs look for when selecting citations
- Direct answer blocks: Paragraphs that lead immediately with the answer, not editorial build-up
- Heading-level relevance: H2s and H3s that closely match likely micro-questions
- Hard statistical data: Specific numbers, dates, percentages the model can extract and quote
- Freshness signals: A clearly visible “last updated” date, ideally within the past 30–60 days for time-sensitive topics
- Clean, structured HTML: Tables, bullet lists, and semantic markup the LLM can parse easily
Three optimisation approaches have produced the most consistent results. First, study your query fan-out — the micro-questions LLMs generate are knowable, and FAQ sections built around them drive measurable citation increases. Second, run embedding similarity analyses to measure how closely your content resembles what’s being cited; this has produced an average 140% increase in AI citations across Brainlabs client tests. Third, treat content freshness as a technical requirement — monthly refreshes of high-demand pages are a minimum for time-sensitive topics.
Measuring AI search performance is hard. Here’s why.
No first-party data from any major AI platform is available at scale. You can’t see actual user queries in Gemini. Without that data, everything is proxies and estimates, which introduces genuine uncertainty into any measurement framework.
The current standard approach: convert keyword data into likely prompts, track them in one of the 30-plus third-party AI tracking tools, and measure brand visibility over time as a proxy for real-world performance. The problem is the data is noisier than it looks. Across models, only 23% of citations are still active after 14 days. There’s only around 45% agreement between platforms on which brand to recommend first. And fewer than 5% of query sets show perfect consensus across all major platforms.
Practical measurement approaches that work now
- Run 100 related prompts per category, four times each — equivalent reliability to running one prompt 400 times, with richer category insight
- Track AI referral traffic in GA4 and run correlation analyses against existing Google search demand data
- Use published AI demand studies (SEOClarity has high quality data) and apply sensible adjustments to existing Google volume estimates
- Document your methodology — any AI traffic estimate will be questioned by leadership, so the reasoning needs to be defensible
Agentic search: the next shift
“We are still diffusing it [agentic search] … and I definitely expect in some of these areas, [20]27 to be an important inflection point.”
Sundar Pichai, CEO, Google
Everything so far sits within the current paradigm, where AI is a layer on top of search. The next shift is bigger. Pichai expects search to transform into an “agent manager” — where AI agents do the research and decision-making on a consumer’s behalf. In the emerging agentic model, an AI agent researches in seconds, filters to a recommendation, and facilitates the transaction, all without the consumer leaving the AI platform. ChatGPT already has Instant Checkout. Google has its Universal Commerce Protocol.
The agents still need to do the research, and they’ll still need information that has to come from somewhere. The question is what content and data sources they’ll trust enough to cite and act on.
Three ways to prepare for agentic search now
- Restructure content for agent consumption. Agents look for explicit conclusions, unambiguous recommendations, precise structured data, and clean HTML. Burying the recommendation in the third paragraph doesn’t work at machine speed.
- Think about your data as an asset agents will want to connect to. B2B API access, consumer-facing authentication, and tiered licensing with AI platforms are all plausible commercial models that don’t require waiting.
- Consider building rather than just feeding. Brands with genuinely valuable expert content have the ingredients to become agentic platforms in their own right — with commerce rails so users can act on recommendations immediately
Where to start
Start with measurement. Set up AI referral tracking in GA4, pick two or three priority categories, and begin tracking 50 to 100 related prompts at low frequency. The data won’t be perfect, but it will be directional.
Next, audit your priority content against the signals LLMs look for: freshness dates, direct-answer formatting, structured data, and heading-level relevance to likely micro-questions. A targeted refresh of high-demand pages applying these principles will tell you whether the changes move citation rates before you scale.
Then build the roadmap for divergence. The overlap between traditional SEO and AI SEO is decreasing. The teams best positioned in two years are the ones building those capabilities now. The window to get ahead of it, rather than chase it, is shorter than it looks.
