Every SEO vendor now sells generative engine optimization: the practice of getting your content cited inside AI-generated answers. The pitch usually arrives with confident tactics and a case study.
Olivier Martinez reviewed 45 studies published between November 2023 and July 2026 and found the evidence does not support that confidence.
What the survey found
The central conclusion, stated plainly in the paper: no reviewed technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability.
The survey separates three claims that vendors tend to blur together.
Evidence is strong that a document already sitting in an engine's context can change how that document gets cited or used. If the model is already reading your page, how you wrote it matters.
Evidence is far weaker that any technique makes a page get retrieved in the first place.
Evidence is close to absent that any of it produces a durable effect on clicks or conversions.
That first gap is where most GEO advice lives. Optimising how you are quoted, given you are already being read, is a different problem from getting read.
One figure is worth keeping: some GEO rewrite methods cut a page's AI retrieval by 16%. The optimisation made things worse.
Why this is hard to measure
Martinez argues GEO is not a single ranking task. It is a stochastic, partially observable pipeline: search activation, crawling, indexing, retrieval, reranking, context allocation, citation, prominence, factual absorption, fidelity, and user behaviour.
You can move one stage and lose at another. You cannot see most of them.
Commercial GEO audits in the survey showed low source overlap between tools, high variability between runs of the same query, and persistent accuracy problems. Ask the same question twice and get different sources.
What we did anyway, and why
We publish llms.txt, structured data on every page, and machine-readable summaries generated from the same source of truth the pages render from. We did that before reading this survey.
We are keeping it, for reasons the survey supports rather than contradicts.
The strong-evidence category covers how a document is used once retrieved. Clean structured data, consistent facts across pages, and dated sources are cheap and they serve human readers regardless. None of it depends on a retrieval claim the evidence does not support.
What we will not do is tell you it drove traffic. Our own Search Console showed five impressions across three months at the time of writing. Any AI-visibility claim we made from that would be invented.
How to read a GEO vendor's pitch
Three questions, drawn from the survey's own critique.
Which stage does this affect? Getting retrieved and being cited well once retrieved are different problems with different evidence bases.
How many runs is the result based on? The survey found high run-to-run variability on identical queries. A result from one check is a coin flip.
Does it end at citations or at clicks? Almost nothing in 45 studies establishes a durable effect on clicks or conversions. A case study showing citations went up has not shown revenue went up.
What holds regardless
The measures that survive the survey's scepticism are the ones that were already good practice.
Publish facts that stay consistent across your own pages. Contradictions between your pricing page and your FAQ get discounted by everything, human and machine.
Date your claims and name your sources. Original data with a method attached is more citable than a rewrite of someone else's post.
Render server-side. In our own crawl of 800 startup sites, 83 served no meaningful HTML until JavaScript executed. No retrieval strategy rescues a page a crawler sees as blank.
None of that is a GEO tactic. It is what a page worth citing looks like.
Sources
- Martinez, O. (2026). Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023-2026). arXiv:2607.14035, submitted 15 July 2026. 45 studies reviewed, November 2023 to July 2026.
- LeadzLander, The State of SaaS Landing Pages (2026). 800 attempted, 675 scored, collected 2026-07-25, seed 20260725.