If you run a business with a website, you have probably had the email by now. Search is dead, SEO is finished, you need GEO โ generative engine optimisation โ and it happens to be what the sender sells.
We went looking for the evidence before recommending it to anyone. What follows is what we found, with every figure linked to its source so you can check it rather than take our word for it.
The short version: the case for paying for GEO in 2026 is much weaker than the pitch suggests, and the thing that should actually concern you is not ChatGPT at all.
The number the pitch leaves out
Here is where search referrals actually came from as of August 2026:
| Source | Share of search referrals |
|---|---|
| 87.63% | |
| Every AI search engine combined | 0.29% |
That 0.29% covers ChatGPT, Gemini, Claude and Perplexity together, and ChatGPT accounts for roughly three quarters of it.
Put plainly: AI search engines currently send about one referral for every three hundred that Google sends. Any proposal that treats GEO as an urgent replacement for search visibility is describing a channel that, for most businesses, barely exists yet.
Two things that complicate the picture
We would be doing exactly what we are criticising if we stopped at the number that suits our argument.
AI referrals convert unusually well. Seer Interactive’s benchmarking found ChatGPT-referred sessions converting at around 15.9%, and Perplexity at 10.5% โ well above typical organic performance. Low volume, but not low value.
It depends heavily on what you sell. A twelve-month study found ChatGPT referrals under-performed Google organic on transactional purchases by about 13%, while over-performing in research-heavy categories โ B2B software, professional services, finance, electronics โ and under-performing in impulse categories like apparel and groceries (summary of the benchmarks).
So if you sell considered, researched services, being present in AI answers is worth something. If you sell impulse purchases, it is currently close to noise.
What the research says about the tactics themselves
This is where it gets uncomfortable for the people selling GEO packages.
In July 2026 a critical academic survey of generative engine optimisation reviewed the field and concluded, in its own words, that “commercial claims about GEO have advanced faster than the evidence.”
Specifics worth knowing before you sign anything:
| The claim | What the research found |
|---|---|
| “GEO lifts visibility up to 40%” | The figure comes from a synthetic test using five pre-supplied documents. It does not describe real-world discovery, and the survey says so explicitly. |
| GEO tactics work broadly | One benchmark found only 3 of 54 tactic-and-domain combinations were significantly positive. |
| One strategy covers every AI engine | Google and Gemini overlapped just 11โ18% at URL level. Optimising for one tells you little about the others. |
| Results are measurable | Ask the same question twice and only 34โ42% of the cited sources overlap. Reliable measurement needs seven to eight repetitions of every prompt. |
llms.txt improves citations |
A study across 300,000 domains found no citation lift, and no AI vendor has committed to reading the file. |
That last one deserves emphasis, because publishing an llms.txt file has
become a standard “we’ve optimised you for AI” deliverable.
Around 8.7% of the top 1,000 websites now publish one.
There is currently no evidence it does anything.
The survey’s most important finding is the least quotable: no reviewed technique showed a stable, repeatable, cross-platform effect on whether content gets discovered in the first place. GEO tactics operate on content the engine has already retrieved. Getting retrieved is still ordinary search visibility.
What does have evidence behind it
Not nothing. Three things held up:
- Topical relevance. Being genuinely, specifically about the thing you are being asked about. The most reproducible factor in the whole literature.
- Extractable evidence โ real figures, clear definitions, quotable sentences โ provided it is truthful and properly attributed.
- Citing external authoritative sources inline. Pages that did this saw meaningful citation gains, and the largest went to sites that were not already dominant: roughly a 115% visibility lift for pages ranking around fifth.
Read that list again. It is indistinguishable from “write specifically and honestly, and show your sources.” There is no new discipline in it. It is what good writing has always looked like, and it is why we have linked every number in this article.
Worth noting what actively backfired: keyword stuffing transfers badly from classical SEO and in some tests reduced AI visibility. The instinct to cram more keywords onto a page is now not merely useless but counterproductive.
The part that should actually worry you
While everyone argued about ChatGPT, the real change happened at Google.
Google made AI Mode the default experience during 2026 and it has since passed a billion users. The consequences for the sites Google used to send traffic to have been severe: publisher clicks down 58%, zero-click searches at 60%, with HubSpot reporting a loss of 70โ80% of its organic traffic and Chegg around 49%.
And the damage is not evenly distributed. Chartbeat’s data shows small publishers lost roughly 60% of their search referrals, against 22% for large publishers.
That is the sentence worth rereading. The contraction is hitting small sites close to three times harder than large ones. If you run a small business website, the threat to your search traffic is not that ChatGPT will replace Google. It is that Google itself now answers the question and keeps the visit.
What we would actually do
We include ourselves in this. Here is the position we have taken on our own site, for whatever it is worth.
Do not buy GEO tooling or a GEO retainer yet. The channel is 0.29% of referrals and the tactics lack evidence. Revisit when that number reaches low single digits โ it is the figure most likely to move.
Do not publish llms.txt. No demonstrated effect.
Stop writing purely informational content. “What is website maintenance” is exactly the query AI answers absorb. It was a reasonable content strategy in 2022 and it is a waste of a page now.
Write for the questions AI cannot close. Someone searching “my contact form stopped sending” needs a person, not a summary. Someone searching for a service provider wants to see the provider. Those clicks still happen.
Use real numbers and cite your sources. The one GEO tactic with support behind it, and it costs nothing beyond doing the reading.
Judge pages by enquiries, not sessions. Traffic becomes a worse proxy for value every year. A page producing four enquiries a month from two hundred visits is working, whatever the graph says.
Do not let search be your only channel by 2028. This is the real conclusion. Referrals, direct demand, reputation and work people can verify all have to grow alongside search, because the share of a search that ends in a visit to your site is being decided by someone else.
The honest summary
GEO is not a scam, and it is not a replacement for SEO. It is a small, promising channel with unusually good conversion, an evidence base that has not caught up with its marketing, and โ for now โ a traffic share that rounds to zero.
The fundamentals did not change. A fast site, genuine expertise, specific writing, published prices and proof you can check are what worked in 2020 and what will still work in 2028. The difference is that fewer people will arrive via a blue link, so the ones who do arrive matter more.
If someone is quoting you for GEO work, ask them for the evidence base. The research is public and free. Their answer will tell you a great deal.