A GEO audit that ends up as a PDF has achieved nothing. It is the first thing I say at kick-off, because the risk is real: produce an impeccable picture of what AI models say about a brand, hand it over, and find six months later that nothing has moved, not because the diagnosis was wrong, but because nobody knew what to do with it. I have described the audit method itself elsewhere. This article takes the problem from the other end: what happens after?
To make it concrete, I am drawing on a representative case, anonymised and with rounded figures: a B2B services SME, roughly fifteen years old, solid SEO, first or second position in Google on its core industry terms. The kind of company with no obvious reason to be invisible. It was invisible nonetheless, in the place where nobody was looking.
A GEO audit is not worth the quality of its diagnosis, but the number of workstreams it opens. The report is not the deliverable: it is the instruction manual for the deliverable.
Before: well ranked, never cited
The baseline measurement is always the same exercise: a panel of prompts, thirty or so in this case, that reproduce the real questions prospects ask, not the keywords from the SEO plan. Not "procurement consultancy Lyon" but "who can help me structure my purchasing function when we go from 30 to 80 employees?". You put those questions to ChatGPT, Perplexity, Gemini and Google's generative answers, you repeat the exercise, and you record three things: does the brand appear, is it cited as a source with a link, and is what is said about it accurate.
First wave result: the brand appears in roughly one answer in twenty, never as a cited source, and of the two occurrences recorded, one describes it with a service scope it dropped six years ago. Meanwhile two competitors, ranked distinctly lower in Google, show up in more than half the answers. This is the typical early-engagement shock, and it is a healthy one: it moves the conversation from "we rank well" to "we rank well in a channel that no longer captures all the demand".
The three causes of the silence
In almost every audit I run, invisibility does not come from some exotic cause. It comes from three defects that compound, and that have nothing to do with ranking.
First cause: the entity is blurry. The model does not know who this brand is. The name comes in four variants depending on the channel, the "About" page tells an emotional story without ever stating what the company does, for whom and since when, and the structured data stops at a breadcrumb. An AI cannot cite what it cannot identify.
Second cause: the content is not extractable. The pages are good: for a human reader who has decided to read. The answer arrives in the fourth paragraph, after some scene-setting, and the claims rest on nothing. Yet the foundational academic study on GEO measures up to roughly 40% additional visibility in generative answers when you add statistics, quotations and sources, without touching the substance. This is not a cosmetic detail, it is the format the machine knows how to slice.
Third cause: nobody else talks about them. This is the heaviest to fix. The brand's entire presence is self-produced: its site, its social accounts, its brochure. No third-party source names it. Yet models rely massively on what others say about you. That is the whole point of the bridge between link building and GEO: the mention weighs more than the link.
What we actually change
Order matters. We start with what depends only on the site, because it is quick and it is the prerequisite: unify the brand name everywhere, rewrite the "About" page so it answers factually "who, what, for whom, since when, where", implement coherent Organization markup, and check that search engine crawlers can reach the pages, including AI crawlers, a topic I cover in the article on llms.txt and controlling what AI reads. Budget one to two weeks.
Then we rework the content, but not all of it: only the pages that answer the prompts in the panel. Each page gains a direct answer up front, dated and sourced figures, subheadings phrased as real questions. In the case described, that meant nine pages rewritten and three created: not a redesign, surgery. Finally, the long workstream: making the brand exist elsewhere. Speaking slots, publishable data, points of view picked up by industry publications. It is the one that produces the most effect and produces it the latest.
After: re-measuring without fooling yourself
This is where most debriefs go off the rails. You ask the questions again three weeks later, the brand shows up twice, and everyone concludes it is working. Except generative engines are structurally unstable at source level. The 2026 measurements converge: over four weeks, only a fraction of cited URLs holds, around a quarter for Google's generative answers, about a third for ChatGPT, a little over 40% for Perplexity, the steadiest of the group. In other words, a good share of citations moves on its own, without you doing anything.
The counterpoint is instructive: Ahrefs research shows that while generative answers rephrase themselves roughly every two days, they almost never change their mind. The meaning stays highly consistent. What moves is the shop window; what holds is the model's judgement on the subject. A serious GEO audit therefore measures the latter, not the former: you do not count "am I showing up today", you count a share of voice averaged across a panel of prompts and several waves. That measurement discipline is what I detail in the article on continuous GEO monitoring.
In the case followed, the trajectory looks like what I see elsewhere. The first two months: almost nothing visible, even though most of the technical work is already done. The third: mentions become accurate. The obsolete service scope disappears from the answers. That is the first real signal, and it is often underrated because it does not push a number up. By the fifth month, share of voice starts pulling clearly away from the noise. This four-to-six-month delay before the effect becomes obvious recurs across most of the field reports published in 2026, and it should be announced at kick-off rather than discovered in a steering committee.
| What we look at | Before the audit | After (month 5-6) |
|---|---|---|
| Google positions on core industry terms | Top 3 | Top 3 (unchanged) |
| Presence in AI answers across the panel | ~5% of prompts | ~35% of prompts |
| Brand cited as a source with a link | Never | Regularly, depending on the engine |
| Accuracy of what is said about the brand | Obsolete scope repeated | Description on point |
| Third-party sources naming the brand | Near zero | Several distinct publications |
| Delay before a clear effect | No | 4 to 6 months |
The most expensive mistake
Judging a GEO audit on a screenshot. An executive asks ChatGPT from their phone, does not see themselves, and concludes the work is useless, or the reverse, sees themselves once and declares victory. Both are wrong in the same way: at this level of volatility, an isolated observation carries no information. There is an added bias few people are aware of: AI models personalise based on history and conversation context. The "home test" run by an executive who has already discussed their own brand with the same assistant is not a test. The only measurement that commits to anything is a protocol: same prompts, same conditions, several waves, averages compared.
What a GEO audit cannot promise
This needs saying as plainly in the sales conversation as in the debrief. You do not control the output of a generative model: you improve the probability of being picked, you do not guarantee it. Nobody can sell a position in a ChatGPT answer the way we used to track a Google position, and anyone promising it is selling something other than GEO. What can be guaranteed, on the other hand, comes down to three points: that the entity is unambiguously identifiable, that the content is in a format the machine can extract, and that the brand exists in sources the models consult. The rest is a matter of repetition and patience.
There is an encouraging corollary. These three workstreams are not a bet on a fad: a clear entity, content that answers, and a brand people talk about also serve classic SEO, conversion and commercial credibility. That is what makes the investment defensible even if generative engines change the rules again, which they will. I connect this logic to the four pillars detailed on my GEO expertise page, and to the full approach behind my GEO audit.
In the crucible: transmuting diagnosis into citation
Between before and after, what changes is not the website: it is what the machine knows how to say about you when you are not in the room. The lead, here, is the audit report filed away in a shared drive: accurate, documented, inert. The gold is the sequence of workstreams it triggers, and the honest re-measurement that says which ones paid off.
A GEO audit is therefore not judged by the thickness of the deliverable, but by the first sentence ChatGPT says about you six months later. If you want to know where you really stand, and above all what would need to change for that sentence to be the right one, that is the purpose of my GEO audit: measure, prioritise, act, re-measure.
Find out what AI says about you, and change it
I measure your share of voice across a panel of real prompts, identify what is blocking (entity, extractability, third-party sources) and hand you a plan of workstreams ranked by priority, with the re-measurement protocol that goes with it.