You have heard the pitch, and it is a seductive one. AI reshuffles the deck. The old signals no longer count. The small player can finally beat the large one, because the model judges the content and not the budget. It is a window, and you should climb through it now.
The measurements say the opposite, and they say it with a depressing consistency.
The problem is not that the window is narrow. The problem is that it opens exactly where it opened before, on the people who are already known.
What follows gives you the numbers, explains why the mechanism is more rigid than the one Google ran on, and isolates the single documented crack in it.
The three tiers
Work published on June 18, 2026 measures brand presence in generative answers at scale: more than one hundred thousand answers, more than one hundred brands, between March and May 2026.
Provenance marking, and it matters: the data comes from Ranqo, a GEO tracking tool vendor. Declarable conflict of interest. To the author’s credit, he concludes himself that causality remains to be established, and he proposes seven verification protocols.
The result is a three-tier ladder, visible from the very first query.
Global brands appear in 73 percent of answers. Mid-tier brands in 44 percent. Niche brands in 11 percent.
Roughly thirty points between each tier. This is not a continuous gradient you climb through effort. These are steps.
A small player does not start from zero, which would at least be a blank page. It starts at eleven percent, and those eleven percent are a ceiling as much as a floor.
Conditional monopoly
A second, independent piece of work measures the underlying mechanism under controlled conditions.
Published on June 16, 2026, it compares products with rigorously identical specifications, identical composition and identical price, querying three commercial models in common use. Provenance marking: arXiv preprint, not peer reviewed, methodology published.
The result: known brands are recommended one hundred percent of the time.
Not sixty percent. Not eighty. One hundred. The authors call it a conditional monopoly: as long as nothing else distinguishes the products, brand recognition decides everything.
This is a laboratory result, with everything that implies about simplification. It also explains the field tiers cleanly.
Why this is more rigid than Google
The natural reflex is to compare with classical search, where large brands were always advantaged. The comparison is instructive, and it is not reassuring.
The advantage large brands held on Google rested largely on inbound links. That was a signal acquired through action: publish, get picked up, get cited. A methodical player could build it over a few years. It was slow and expensive, but it was a path.
Here the dominant signal is not a link graph the engine recomputes. It is a learned distribution over a training corpus in which large brands are massively overrepresented, topped up by a retrieval stage that draws on third-party sources where they are overrepresented just as heavily.
An analysis covering 75,000 brands gives the measure of the shift: web brand mentions correlate at 0.664 with visibility in generated summaries, against 0.218 for backlinks. Provenance marking: SEO tool vendor, self-reported data, and the vendor takes care to write that correlation is not causation and that the relationship is probably confounded by brand size.
The top three factors identified are all off-site. And one secondary figure from the same analysis deserves quoting: 26 percent of brands have no mention at all in those summaries.
The lever changed nature. A link was acquired. A mention is earned or bought in brand awareness, which is not the same budget line.
What argues the other way
Before we conclude, two elements prevent any claim of total lock-in, and leaving them out would be dishonest.
Generative engines cite more diverse sources than Google. Work from December 2025 covering 55,936 queries, comparing six generative engines against two traditional ones, establishes that the former cite 37 percent unique domains, a higher diversity than Google and Bing. Provenance marking: arXiv preprint, not peer reviewed. The same work adds that they do no better on credibility or political neutrality, but on the narrow question of concentration, the lock-in story needs qualifying.
There is one measured crack, and only one. The conditional monopoly study tested what breaks the effect. The answer: an advantage of 0.1 stars on the average rating is enough to break the preference for the known brand.
One tenth of a star. Not a site rebuild, not a configuration file, not a rewritten product sheet. A genuinely better customer rating.
What this changes for a small business
Here we set the pitch against the evidence, line by line.
| What you are sold | What the studies measure |
|---|---|
| “AI judges the content, not the brand” | At identical specifications, the known brand wins 100 percent of the time |
| “It is a window of opportunity for small players” | Niche tier at 11 percent, thirty points below the mid tier |
| “Optimize your site to get cited” | The top three visibility factors are all off-site |
| “Backlinks no longer count” | Correct, but what replaces them correlates three times more strongly and is harder to acquire |
| “GEO democratizes visibility” | No measurement supports it. One crack exists: the customer rating |
The strategic conclusion is not to give up. It is to stop spending on the weakest lever.
If you are a niche brand, the spending with a chance of moving your tier is not the optimization of your pages. It is whatever produces mentions elsewhere: presence on the comparison sites in your sector, the trade publications, the communities where your category gets discussed. A study run across several verticals and several languages, published as an arXiv preprint, confirms a systematic bias toward exactly those authoritative third-party sources, at the expense of content you publish yourself.
And it is whatever improves the rating. One tenth of a star is the only item on that list an established competitor cannot take back from you in three weeks.
What you do tomorrow morning
Locate your tier honestly. Ask your ten category queries on every platform, several times each, varying the wording, and count the share of answers where your name appears. A single measurement is worth nothing, because generative visibility is a distribution rather than a point.
If you are somewhere around ten percent, you are in the niche tier, and no on-page optimization budget will carry you thirty points.
Reallocate the spending across three lines instead: presence in the third-party sources of your sector, real coverage of your subject rather than its formatting, and the customer rating.
GEO does not reshuffle the deck. It redistributes the same advantage under a new name, and makes it harder to work on than it was before. The one thing worth holding against it is that it gets sold as the opposite.
Sources
- Kumar, P. (2026). Generative Engine Optimization at Scale: Measuring Brand Visibility Across AI Search Engines, arXiv:2606.20065
- Chu, X. & Hou, Y. (2026). Incumbent Advantage: Brand Bias and Cognitive Manipulation Dynamics in LLM Recommendation Systems, arXiv:2606.17443
- Ahrefs (2025). An Analysis of AI Overview Brand Visibility Factors (75K Brands Studied)
- Zhang, P., Ye, Q., Peng, Z., Garimella, K. & Tyson, G. (2025). Source Coverage and Citation Bias in LLM-based vs. Traditional Search Engines, arXiv:2512.09483
- Chen, M., Wang, X., Chen, K. & Koudas, N. (2025). Generative Engine Optimization: How to Dominate AI Search, arXiv:2509.08919
- Schulte, J., Bleeker, M. & Kaufmann, P. (2026). Don’t Measure Once, arXiv:2604.07585
<strong>LaFactory</strong> measures AI visibility with a published protocol: repeated measurements, paraphrases, control group. No guaranteed placement, ever. Contact us to scope an audit.