Your GEO budget goes to your site. Page rewrites, paragraph restructuring, schema markup, a configuration file at the root, perhaps a technical audit on top. That is logical, because your site is the only thing you control and the only thing an agency can deliver against a fixed fee.
It is also the place where the measurements report the weakest effect.
The problem is not that your site is bad. The problem is that generative engines go looking elsewhere, systematically, and that elsewhere is not you.
What follows gives the size of the effect, corrects a statistic that almost everybody quotes wrongly, and proposes a different allocation of the same budget.
The bias toward third-party sources
The most direct work on the question was published on September 10, 2025 by a team at the University of Toronto. Its protocol: large-scale controlled experiments, across several verticals, in several languages, with paraphrased queries and a comparison between generative search and Google.
Provenance marking: arXiv preprint, not peer reviewed, methodology described.
Its central finding is a systematic and pronounced bias toward earned media, meaning authoritative third-party sources, at the expense of the content you publish yourself and of your social channels. The contrast with Google is sharp: the classical engine produces a noticeably more balanced mix of owned, earned and social.
The same work documents two secondary properties that matter for anyone operating outside English: a marked advantage for large brands, and instability by phrasing and by language. What is measured in English does not transfer mechanically to any other market.
The statistic everybody quotes wrongly
A ranking of the domains most cited by AI circulates everywhere: Reddit 40.1 percent, Wikipedia 26.3 percent, YouTube 23.5 percent. Those numbers are real. They come from a July 2025 analysis of 5,000 randomly drawn keywords and more than 150,000 unique citations, across Google, generated summaries, Google’s conversational mode, ChatGPT and Perplexity. Provenance marking: SEO tool vendor, self-reported, methodology partially published.
They are almost always presented as shares of citations. They are not.
The proof is arithmetic and takes ten seconds. Add up the first twenty lines of the ranking and you pass 280 percent. Shares of a single total cannot add to 280 percent.
What that column actually measures is a frequency of appearance: the percentage of queries for which the domain shows up at least once, across all engines combined. That is a different thing, and the gap is not cosmetic. Saying that Reddit captures 40 percent of citations is false. Saying that Reddit appears in 40 percent of answers is true, and already considerable.
We published that error ourselves before going back to the source, which is why it is worth flagging plainly. That is exactly the mechanism this series documents: an accurate number, a misread metric, and a chain of repetition that never walks back to the original.
Read correctly, the figure is still eloquent. In Google’s conversational mode, Reddit, YouTube and Facebook appear in more than 68 percent of results that carry additional links.
The other half of the problem: what does not work on your site
The third-party bias would matter less if on-page work produced at least a measurable effect. That is not what the widest available measurement establishes.
A study from May 2026 covering 252,000 trials, six models and a factorial design testing eighteen content factors, with brand anonymization and counterbalanced source ordering, reaches an unambiguous conclusion: pure formatting edits have almost no effect. Structure, lists, chunking, markup.
Provenance marking: team at a marketing software vendor, declarable conflict of interest, protocol published.
Your budget is therefore aimed at the wrong place and at the wrong lever at the same time.
What contradicts this
Two results prevent any conclusion that on-page work is useless everywhere.
Where your page is itself the third-party source, optimization works. An MIT testbed built on 13,747 product queries crossed with Amazon listings, published as an arXiv preprint, shows real gains, and specifies that under a simple defense those gains correspond to a genuine improvement of the content. The difference is contextual: a product listing on a marketplace is not your site, it is already a third-party source from the engine’s point of view.
Concentration is not total. Work from December 2025 covering 55,936 queries, comparing six generative engines against two traditional ones and published as an arXiv preprint, measures higher source diversity among the former, at 37 percent unique domains. The story of a market locked onto five domains is therefore overstated, even if the appearance frequency of the large platforms remains overwhelming.
Reallocating the budget
Here we put every line item next to what the evidence supports.
| Line item | What it weighs today in a typical GEO offer | What the measurements justify |
|---|---|---|
| Page rewriting and restructuring | Most of the fee | Near-zero effect for pure formatting |
| Markup and configuration file | A recurring line | No demonstrated effect on citation |
| Real coverage of the subject, in its own vocabulary | Rarely isolated | Leading determinant, alongside position |
| Presence on comparison sites and trade publications | Almost never included | Aligned with the measured earned media bias |
| Presence in the communities where the category is discussed | Almost never included | Aligned with the observed appearance frequency |
| Listings on marketplaces and specialized directories | Treated as e-commerce | Demonstrated effect, MIT testbed |
| Customer ratings and reviews | Out of GEO scope | One tenth of a star breaks brand preference |
The first three lines are the bulk of what gets sold. The next four are the bulk of what gets measured.
The objection you are going to hear
Your vendor will answer, fairly, that off-site work is slow, hard to guarantee, and impossible to bill cleanly by the month. That is true.
They may add that this is public relations rather than GEO. That is also true, and it is precisely the point. Google’s own documentation states that optimizing for generative search features remains search engine optimization.
The question is not what to call the work. The question is whether you are paying for the strong lever or the weak one.
What you do tomorrow morning
Take an inventory, across your ten category queries, of what the engines actually cite. Not your share of voice: the list of sources. Ask each question several times, vary the wording, and record the domains that keep coming back.
You will get a short list, often around fifteen domains: sector comparison sites, two or three trade publications, a forum, a video platform, sometimes a directory. That is your real target.
Then count how many of those domains mention you today, and with what information. You will regularly find an obsolete price or a discontinued offer sitting on a comparison page you had forgotten existed.
Then allocate your budget in this order: correct the information on the sources that already cite you, obtain a presence on the ones that do not, and only after that improve your own pages.
Your site remains necessary. It is simply the place where the marginal effort pays least, and it is the place you are being told to concentrate all of it.
Sources
- Chen, M., Wang, X., Chen, K. & Koudas, N. (2025). Generative Engine Optimization: How to Dominate AI Search, arXiv:2509.08919
- Semrush (2025). AI Mode comparison study : 5 000 mots-clés, 150 000+ citations
- Vishwakarma, R., Kumar, S. & Jamidar, R. (2026). What Gets Cited: Competitive GEO in AI Answer Engines, arXiv:2605.25517
- Zhang, P. et al. (2025). Source Coverage and Citation Bias in LLM-based vs. Traditional Search Engines, arXiv:2512.09483
- Bagga, P. et al. (2025). E-GEO: A Testbed for Generative Engine Optimization in E-Commerce, arXiv:2511.20867
- Chu, X. & Hou, Y. (2026). Incumbent Advantage, arXiv:2606.17443
- Google Search Central. Optimizing your website for generative AI features on Google Search
<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.