You can be the number one result and still be missing from the answer

Type “best health insurance india” into Google and look carefully at what comes back. There’s an AI Overview sitting at the top, four or five confident sentences with a row of small source chips attached to it. Underneath that sits the organic result that ranks number one. Now check whether that number one result is one of the sources the Overview pulled from.

Very often, it isn’t.

That gap is the reason my co-founder Sankalp and I started Geology. It came out of a specific, repeatable thing we kept seeing on real queries for real clients: a brand had done everything the SEO playbook asked of it, won the position it was told to win, and still had no presence in the sentence the buyer actually read.

Search didn’t shrink. It got absorbed.

In February 2024, Gartner predicted that traditional search engine volume would fall 25% by 2026 as people moved to chatbots. It’s 2026 now, so we can mark the homework. That’s not really what happened. People didn’t abandon the search box in the numbers anyone forecast, and Google didn’t lose the market to a chat interface.

What happened instead is stranger and, for anyone whose business depends on being found, more consequential. Google took the chatbot and put it inside search. At I/O in May 2026, Sundar Pichai said AI Overviews had passed 2.5 billion monthly active users, and AI Mode had crossed 1 billion in about a year. The query volume stayed. The interface on top of it changed underneath everyone.

So the doom prediction was wrong, and I’m glad it was wrong. I’m an optimist about this stuff. But getting the collapse wrong made a lot of people relax about the wrong thing. They watched for traffic to fall, saw it hold, and concluded that nothing underneath had changed.

Pew Research put numbers on what did change. In July 2025 it tracked the browsing behaviour of 900 US adults across 68,879 Google searches. When an AI summary appeared, users clicked a traditional search result in 8% of visits, against 15% when no summary appeared. Roughly half. The links inside the summary itself got clicked on 1% of visits. Users also ended their session entirely on 26% of pages carrying a summary, against 16% without one.

Read that 1% again, because it’s the number that reorganised how I think about this work. Being cited in an AI answer almost never sends you a visitor. The citation is the value. Someone asked a question about your category, and the answer they received either had you in it or didn’t.

Ranking and being cited are two different jobs

Here’s the thing though. Everyone accepts the above in the abstract and then goes back to optimising for position, because position is what the tools measure and what the reporting line expects.

The mechanics underneath are different. I wrote a while back about how ChatGPT decides whether to search at all, and about a behaviour I still find remarkable: for a meaty comparison query, it pre-selects a handful of brands from its own knowledge before it runs a single search. In the example I looked at, it had settled on Nordstrom, Macy’s and DKNY before it fetched anything. A model doing that is assembling a shortlist from what it already believes about your category, then going looking for material to support the answer. The ranked list barely enters into it.

Two things follow from that. The first is that you can hold position one and never enter the shortlist, because the model drew that shortlist from training data and brand-level familiarity, long before it looked at today’s SERP. The second is that a page can rank beautifully for humans and be useless as source material, because it renders its actual content in client-side JavaScript, or buries the number a model needs under a quote flow, or answers the question in a video.

Traditional SEO tools will tell you that you rank third. They won’t tell you the model has never once used you.

Finance and insurance is where the gap gets expensive

I spend most of my research time on insurance and financial services, partly because it’s the category I know best and partly because it’s where this problem stops being a marketing inconvenience.

In June 2026 we ran a study of 100 US insurance buyer prompts across seven insurance types, through ChatGPT, Perplexity, Gemini and Google AI Overviews, tracking fifteen insurers. Some of what came back was predictable. Incumbents dominate: they take between 66% and 82% of brand mentions depending on the engine, and challengers scrape 13% to 16%. If you’re a large carrier, that sounds like good news, and up to a point it is.

Then look at where the engines get their material. In ChatGPT’s answers, carrier domains accounted for 1% of citations. Aggregators accounted for 47%. Perplexity cited Reddit in 75 of the 100 answers we ran. In Google AI Overviews, aggregators took 30% of citations against 22% for carriers.

One encouraging result sits in the same data, and it cuts against the incumbent story. Narrow the question and the hierarchy breaks. Next Insurance took 50 mentions across small-business prompts, and Lemonade led both renters and pet insurance, categories where none of the big carriers had built the same density of specific, retrievable material. Overall share follows brand size. Category share follows whoever wrote the clearest thing about that category, which is a fight a smaller carrier can pick and win.

So the brand gets mentioned, and somebody else supplies the facts behind the mention. State Farm appears in a third of answers about insurance, and the engine builds its description of what State Farm covers and what it costs out of a comparison site’s summary of State Farm. The carrier is present as a name and absent as a source.

In a category selling shoes, that’s a nuisance. In one selling indemnity, it’s a different kind of problem. We’ve watched engines quote an expired promotional rate, an intro fee presented as the standard fee, and a premium a carrier revised two quarters ago, all attributed confidently to the brand, with nobody inside that brand having reviewed a word of it. Financial promotions rules are technology-neutral. A regulator asking who approved that description of your product will not be satisfied by an explanation involving retrieval-augmented generation.

The uncomfortable part is where this comes from. Most of these carriers have accurate, current, compliance-approved numbers on their own websites. The numbers sit behind a quote flow that renders client-side, so the crawler arrives, finds an empty shell where the premium should be, and fills the gap from the most legible source available, which is a marketplace page from last year.

What we actually decided to measure

Once you accept that citation and ranking are separate, the reporting problem becomes obvious. Every tool in the SEO category answers “where do I rank.” Almost none of them answer “did the model use me.”

That’s the question we built Geology around. We run structured prompt sets against the engines the way you’d run a survey, at volume and on a schedule, and we count share of voice, which brands get named, and share of citation, whose pages the answer got built from. The interesting client conversations start when those two diverge. A carrier with high mention share and near-zero citation share has a specific, fixable problem: the engines are talking about them using someone else’s description. That’s a different remedy from “publish more blog posts,” and you can only see it if you’re measuring the right thing.

Sankalp came at this from product design and spent his time on the measurement side, working out how you observe brand presence inside an answer that’s generated fresh every time and never looks identical twice. I came at it from the search and insurance side. We started out trying to answer a question about a health insurance query, found that the existing tools couldn’t answer it, and built the thing that could. The agency came after.

The optimistic reading

None of this reads to me as bad news, and I’d rather not add to the pile of writing that treats AI search as a thing being done to publishers.

The old game rewarded whoever could best reverse-engineer a ranking algorithm in a given quarter. It was a tax on cleverness and it reset every core update. What replaces it rewards being the clearest, most current, most machine-legible source of truth about your own product. Publish your real rates in text a crawler can read, and answer the comparison questions your category gets asked rather than the ones your brand guidelines prefer. Make it easier for a model to quote you than to quote a page about you.

Strip the jargon off and that is just a request to be useful and unambiguous about what you sell. I’ve spent enough years in SEO to find it a relief that the winning move has quietly become “be the best available source” rather than “guess this quarter’s weighting.”

The carriers that figure this out get one thing worth having: the version of themselves that a billion people’s assistant repeats back to them, in their own words, with their own numbers, checked by their own compliance team. The clicks are gone either way.

Right now, for most of them, someone else is writing that sentence.