In April 2025, Ahrefs published a number that made a lot of people uncomfortable: pages with an AI Overview above them lost about a third of their clicks. Ten months later they re-ran it with two more years of Search Console data. The number had gone from 34.5% to 58%.
Not 58% of all traffic. 58% of the clicks a number-one ranking used to earn. You still rank. The page still exists. The click just doesn't arrive.
100 clicks a top-ranking page used to get
Method: 150,000 keywords with AI Overviews against 150,000 without, December 2023 vs December 2025, aggregated Google Search Console CTR. Ahrefs is careful to note the non-AIO control group fell too — this is an acceleration of zero-click search, not a brand-new disease.
If you sell websites for a living, that number is an existential problem. If you own a plumbing company, it's something stranger: the customer still exists, still has the emergency, still hires somebody. They just decide before they ever land on a page. Which means the whole question has changed from how do I rank to how do I get named.
That discipline picked up a name — generative engine optimization, GEO, sometimes answer engine optimization — and in 2026 it stopped being a thing agencies invented to have something new to sell. Here are the five shifts that actually changed this year, each one with the study behind it, and then the part that matters for a local business.
Shift 01 — The vocabulary stopped being optional
Search demand is the cheapest proxy for what an industry is worried about, and the curve here is not subtle. Ahrefs' 2026 trends analysis tracked the clusters:
Search demand growth, 18 months
generative engine optimization18 MONTHSgeo vs seo18 MONTHSzero-click search strategy18 MONTHSai search trackingYEAR ON YEARThe tell isn't the growth, it's which terms grew. "GEO vs SEO" rising almost as fast as "GEO" itself means the market is still arguing about definitions — which is exactly the moment when nobody has locked in an advantage yet.
Take the practical reading from that: your competitors are googling what this is. Very few of them have done anything about it.
Shift 02 — Being talked about beat being linked to
This is the one that should reorganise your priorities. Ahrefs took 75,000 brands and correlated dozens of measurable signals against how often each brand appeared in ChatGPT, Google AI Mode and AI Overviews. Backlinks — twenty years of SEO's favourite currency — finished near the bottom.
What correlates with being named by AI
SOURCE: AHREFS — 75,000 BRANDS
Spearman correlation with AI visibility. Correlation is not causation — starting a YouTube channel tomorrow does not summon citations. What it says is that brands AI names are brands the web talks about, in plain sentences, in a lot of places.
For a home service company this is unusually good news, because you cannot buy your way to 0.74 but you can earn your way there cheaply: the review you ask for, the neighbourhood Facebook thread, the supplier's "installers we recommend" page, the local news piece about the storm you worked through. Every one of those is an unlinked mention, and unlinked mentions are what the data points at.
Shift 03 — Each engine reads a different library
People say "AI search" as if it were one place. It isn't. The systems draw on visibly different source pools, which is why a business can be quoted constantly in one and never appear in another.
Where the answers come from
CITATION ANALYSES, 2026 — SEE SOURCES BELOW
CHATGPT
Wikipedia
~47.9% of cited sources
Leans on encyclopaedic and established reference pages. Hard to game, slow to change, rewards being a well-documented entity.
PERPLEXITY
~46.5% of citations
Heavily weighted to forums and recent posts. Real people describing real experiences — including the thread where somebody asks who to call in your town.
GEMINI
Almost never Reddit
~0.1% of responses
A different pool again, closer to Google's own index. Same question, same city, different three companies in the answer.
Across a large-scale analysis of AI citations in early 2026, only about 11% of cited domains appeared in both ChatGPT and Perplexity. Winning one is not winning the category.
The operational lesson is boring and correct: check all of them, and treat a strong result in one as a sample of one. Our 10-prompt check exists for exactly this — same ten questions, four surfaces, write down who got named.
Shift 04 — llms.txt turned out to be a dead end
Through 2025 the advice everywhere was to publish an llms.txt file: a tidy markdown summary of your site, for the models. It sounded right. It cost an afternoon. A quarter of the web did it.
Then somebody measured. Ahrefs looked at every domain in their Web Analytics that received traffic in May 2026 — 137,210 of them — and checked who was actually fetching those files.
llms.txt, measured
SITES THAT PUBLISHED ONE
28%More than one in four domains shipped an llms.txt file on the strength of speculation alone.
OF THOSE, GOT ZERO REQUESTS
97%No AI bot, no human, nothing fetched the file in the entire month. All the traffic went to the other 3%.
Google has separately documented that llms.txt has no effect on Search or AI Overviews. No major AI provider has committed to reading it in production. It is harmless — it is just not a strategy.
Worth sitting with, because it is the shape of most GEO advice right now: plausible, cheap, widely repeated, and untested. The counter-move isn't cynicism, it's asking for the measurement before you spend the afternoon.
⚠ STILL NOT REAL, STILL BEING SOLD
"Guaranteed #1 in ChatGPT." There is no ranking to buy, no ad slot, no placement. "AI indexing / registration service." There is nothing to register with. "Results in 7 days." Sources have to be re-crawled and re-ingested before a change reaches an answer — four to eight weeks is the honest range. If a pitch contains any of the three, the rest of it is not worth reading.
Shift 05 — Your website now has more machine visitors than human ones
In June 2026, Cloudflare's CEO noted that bot traffic had crossed over human traffic for the first time. Meanwhile "agentic SEO" as a search term grew by nearly 6,000%, off a tiny base, because everyone can feel the same thing coming: the next visitor to your site may not be a person at all, but something reading on a person's behalf and reporting back three options.
That reframes a lot of web design advice. A hero video does nothing for a reader that cannot watch. A phone number rendered inside an image does not exist. A price hidden behind "contact us" cannot be compared, and a thing that cannot be compared does not make the shortlist.
The part that's actually about you
Everything above is industry-wide. Here's the local cut, and it's the most interesting number in this article. SOCi's 2026 Local Visibility Index measured more than 350,000 business locations across both surfaces:
Same businesses, two worlds
SOURCE: SOCi LOCAL VISIBILITY INDEX 2026 — 350,000+ LOCATIONS
VISIBLE IN GOOGLE'S LOCAL 3-PACK
35.9%The map result you already work for. A third of locations make it.
RECOMMENDED BY CHATGPT
1.2%The same businesses, asked the same kind of question, in the place customers increasingly ask it.
Roughly thirty times more selective. Read it twice, because it cuts both ways: your Google ranking does not carry over — and almost nobody in your city is competing for the other list yet.
That second half is the whole opportunity. Getting into the local 3-pack means beating everyone who has been optimising for a decade. Getting named by an assistant means being the one company in town whose information is clean, specific and corroborated. In most American cities, that bar is on the floor.
Score yourself in sixty seconds
Tick what's true today. No email, nothing sent anywhere — the number updates as you go.
0/100
There isn't enough consistent, quotable material for an assistant to risk naming you. Fixable — most of your competitors are here too.
Whatever you scored, the honest version of that check is not a checkbox — it's your actual homepage, read the way a crawler reads it. Run the free scan and you'll get the same categories graded against the real HTML, including whether your robots.txt is quietly blocking the bots.
What to do in the next thirty days
Ordered by what the 2026 data says moves the needle, not by what's easiest to sell.
Notice what isn't on the list: llms.txt, "AI-optimised" content spinners, schema stuffed with fields you can't verify, and anything sold with the word guaranteed. Notice also that six of the seven are things a business owner can do without a developer — which is the actual state of GEO in 2026. The bar is low, the tactics are unglamorous, and the window is open precisely because neither of those facts is exciting enough for most people to act on.
The one-sentence version
- The click is being replaced by the mention — plan for the mention
- Being described in plain sentences across the web beats accumulating links
- Every engine reads a different library, so check all four before you conclude anything
- Ignore tactics nobody has measured, including the popular ones
- Locally, 1.2% versus 35.9% is not a warning — it's an unclaimed lane
I since ran these checks across 210 real contractor websites and published everything I found — the numbers are here, including the one that surprised me most.
If you want the shortest possible starting move: scan your homepage, fix whatever comes back red, then run the ten prompts and put a reminder in your calendar for thirty days out. That's the whole programme. Everything else is detail.