AI Search Is Changing SEO: What To Do Now

AI Overviews now appear on more than 40% of Google queries. AI-generated local packs surface roughly 68% fewer businesses than a classic three-pack. Around 41% of small and mid-size businesses report having adopted some generative-optimisation practice.

Those three numbers describe a real change, and most of the commentary about them is either panic or a rebrand. What follows is the mechanical version: what actually differs, what genuinely transfers from conventional SEO, and where the honest limits of measurement sit.

The one mechanical difference that matters

A traditional search engine ranks a list of documents and lets you choose. A generative system retrieves candidate passages, synthesises them into a single answer, and cites a handful of sources.

That is the whole thing, and everything downstream follows from it. The unit of competition stops being the page and becomes the extractable passage. You can rank fourth organically and be the sentence the model quotes. You can rank first and be paraphrased without attribution.

It also produces a genuinely counterintuitive result: page-level authority matters less here than it does organically. A precisely worded, self-contained, factually specific paragraph on a modest domain is frequently a better extraction candidate than a hedged paragraph on an authoritative one. Smaller sites get cited over larger ones more often than the classic ranking model would predict.

Three things that have to be true before anything else

Before passage-level work is worth doing, three foundations have to hold. In order:

1. Retrievability

Can a system that reads your HTML without executing JavaScript find your content? A meaningful share of AI crawling does exactly that, and does not wait around for a client-rendered page to hydrate.

This is the single highest-value fifteen minutes available: fetch your own page with JavaScript disabled and read what comes back. If a meaningful portion of your body content is missing, you are invisible to that channel regardless of how well written it is. The site looks perfect to every human who visits, so nobody suspects it — and conventional audit tools that crawl rendered output will not flag it either.

This is disproportionately common in markets with strong engineering cultures. California sites are the most likely we see to have this problem, because the technology sector’s preference for client-rendered applications propagates into every category there.

2. Entity clarity

A model cannot confidently cite an entity it cannot resolve. If your business is described four different ways across your site, your schema, your Google profile and three directories, a system trying to answer a question about you has to reconcile that — and when it cannot, it either declines to cite you or cites whichever version is most prominent, which is frequently the outdated one.

The fix is tedious rather than difficult: one canonical description used verbatim across your site and your Organization markup, sameAs links to your real profiles, and reconciliation of third-party sources carrying the wrong version.

3. Machine-readable structure

Correct FAQPage, Service, Organization and BreadcrumbList markup gives a retrieval system a pre-parsed, unambiguous version of your page. It materially improves the odds that the fact extracted is the fact you intended.

This is also, remarkably, still a competitive gap. When we audited the five agencies ranking hardest in our own home market, not one shipped FAQPage schema despite having FAQ content on the page. One marked a service page up as an editorial Article. One leaked LegalService markup onto its own homepage from a client subsite. Being simply correct beats most of that field.

Writing for extraction

Once the foundations hold, the content work is specific and unglamorous.

Answer first, qualify second. Inverted pyramid. A passage that opens with three sentences of context before reaching the answer does not survive being lifted out of that context, so it does not get lifted.

Be specific enough to be worth quoting. This is the rule that does most of the work. Compare:

We deliver outstanding results for our clients.

An 84% organic traffic increase over three months for a single-location client.

The first is unattributable — there is nothing in it a model can verify, contrast or usefully quote. The second is a discrete, checkable claim with a figure and a timeframe. A system choosing which of several sources to name uses specificity as its main signal that a passage contains information rather than positioning.

The same applies to conditions and caveats. “Local SEO helps you reach nearby customers” is vague. “Proximity decay typically limits local pack range to two or three miles in dense consumer categories” is citable.

Date and attribute. An undated statistic is far less citable than the same statistic with a source and a year.

Get facts out of images. Pricing in a PDF, comparisons rendered as graphics, key data inside an infographic — all invisible. Publish the substance as text alongside the visual asset. This is especially acute for industrial and technical businesses, where the most commercially valuable content is routinely locked in downloadable specifications.

The measurement problem, stated honestly

There is no Search Console for ChatGPT. Anyone selling you a position number for a generative surface is describing something that does not exist.

What can be measured is presence and share:

• A fixed prompt set reflecting how your buyers actually ask, run monthly across the major assistants and AI Overviews, recording who is cited for each.
• AI Overview appearance rate for your tracked keywords, visible in Search Console impression patterns.
• Assistant referral traffic, where those surfaces pass a referrer.
• Branded search volume, which reliably lifts when a business starts being named in AI answers and is frequently the earliest observable signal.

That is a real measurement programme. It is a different instrument than a rank tracker, and the trend line is what matters rather than any single reading.

What this changes about traffic, and what to do about it

The pattern worth watching for: stable rankings, falling clicks. Compare impressions against clicks per query in Search Console. If impressions hold while clicks decay, you are still surfacing but something above you is satisfying the query.

It concentrates almost entirely on informational searches, which is diagnostic in itself. If your how-to and what-is queries are collapsing while your commercial and branded queries are unaffected, that is zero-click erosion rather than a ranking problem.

The strategic response is uncomfortable but straightforward: stop treating informational content as a traffic channel and start treating it as a citation and authority channel, while rebalancing conversion emphasis onto the commercial-intent queries where a summary cannot substitute for a provider. That is a genuine repositioning of what your content is for, not a tactical fix.

Is it too early, or too late?

Both, uncomfortably. Around 41% of SMBs have already adopted some generative practice, so the first-mover window is narrower than it was a year ago.

We would rather correct our own earlier assumption here than repeat it. Broad market research suggested GEO was wide open in the agency space — and for most agencies in our region it is. But a deeper audit of who actually ranks for our HQ market found a direct competitor already running three distinct, well-built GEO service pages at roughly three thousand words each. That is a genuine, well-resourced content silo. We do not tell clients this space is empty.

What makes the investment defensible now rather than speculative is that the foundational work overlaps heavily with good conventional SEO. Retrievable content, clean entity data, correct schema and answer-first structure all improve classic organic performance too. You are not betting the budget on a single surface.

Where to start this week

1. Fetch your homepage and two key service pages with JavaScript disabled. Read what comes back.
2. Check whether you ship FAQPage schema, and whether it matches your visible content exactly.
3. Ask an assistant the question your best customer would ask. See who gets named.
4. Look at your robots.txt and decide, deliberately, whether AI crawlers should be there.

That is a morning’s work and it will tell you more about your position in this channel than any amount of reading about it.

If you want the fuller version, our AI SEO service starts with a citation baseline across your buyers’ real questions — and we will tell you plainly if your problem turns out to be technical rather than generative, which it frequently is.