What to do when a client isn't improving their AI visibility: diagnosis and the hard conversation
Month four of the service. The appearances chart is still flat, the client starts asking questions in a different tone and you have two options: dress up the report or sit down and diagnose properly. This article is for the second.
Worth saying plainly: this will happen to you. AI visibility — GEO (Generative Engine Optimisation), if you want the technical name: measuring and improving how a brand shows up in the answers given by ChatGPT, Gemini, Perplexity and Claude — is work on sources and authority that in some clients takes longer than their patience can bear, and in some cases the diagnosis will reveal that the problem is the approach, not the execution. The difference between a serious agency and one that sells smoke isn't that the serious one always succeeds: it's what it does when it doesn't.
Before the tree: confirm it really isn't improving
Three quick checks, because sometimes the problem is measurement:
- Are you measuring the same way each time? Same prompts, same frequency, all 4 AIs. If you've changed the list along the way or you only look at ChatGPT, you don't have a comparable series — you have anecdotes.
- Are you only looking at appearances? There are intermediate improvements that arrive first: false data corrected, the client's website cited as a source for the first time, a better description in the brand prompts. If that's moving, the system is working even if the appearances in discovery haven't arrived.
- Is it flat or is it volatile? AI answers swing week to week. One bad month after two good ones isn't stagnation; six weeks with no positive signal in any AI is.
If after this the conclusion is still "not improving", on to the tree.
The diagnostic tree: four branches
Branch 1: a sources problem? The AI can't recommend what it can't find. Symptom: it's not that the client shows up badly — it's that they don't exist in the answers, and when you ask about them directly the AI has little or nothing to say. Common cause: no reviews or very few, outdated or contradictory directories, zero third-party mentions (press, rankings, forums). It's the most common branch in SMEs and the most treatable: the plan becomes offsite — listings, reviews, mentions — for one or two quarters.
Branch 2: a content problem? The external sources exist but the client's website never appears among the ones cited. Symptom: in Perplexity and Gemini you see that the answers in their category cite blogs, comparisons and media — and their website never. Cause: a website that doesn't answer questions (it only describes itself), with no clear or structured data, no citable content. The plan: content that answers exactly the prompts you want to enter, and structured data.
Branch 3: an authority problem? The work is done — sources up to date, content published — but the answers are always monopolised by the same three or five big brands. Symptom: an almost identical list of recommendations across all the AIs and all the generic prompts. Here the honest answer is twofold: short term, move the fight to more specific, niche prompts where there is a gap; long term, build the missing authority, knowing it's the slowest branch. It's also a warning about expectations: the starting point varies enormously by sector, and the goals must be set accordingly.
Branch 4: a time problem? Everything above is reasonably fine and the work is recent: the corrections are three weeks old, the content a month. The AIs take time to incorporate new sources, and that lag nobody controls. Symptom: there's nothing broken to fix. Plan: verifiable intermediate milestones (the source is corrected, the content is indexed) and an agreed review date.
The branches combine — weak sources and a sector with concentrated authority is the usual cocktail for long stagnation — but it forces your team to name the main branch. "It's just that this is slow" isn't a diagnosis; it's branch 4 without having ruled out the other three.
The honest conversation: the script
Ask for the meeting yourself before the client asks for it. Getting to the bad news first is half the credibility.
1. The data with no make-up (first five minutes). "I've asked for this meeting because the visibility results aren't progressing at the pace we expected, and I want to show you exactly where we are and what we've found." And the real chart, flat, on screen. No cushioning.
2. The diagnosis. "We've analysed why, and the main cause is this" — and the branch, in plain language. For branch 1: "the AIs build their answers from reviews, directories and mentions, and right now you barely exist there; until that base is built, there's no raw material to recommend you with."
3. What changes. The adjusted plan, concrete: what stops, what starts, what intermediate milestone they'll see in 30-60 days. If you committed to work KPIs and not results KPIs — as you should, here's which and which not — this is the moment to show that what was committed is done and what's being readjusted is the strategy.
4. The decision, shared. "I propose we continue with this plan for three months with these control signals. If you'd rather pause and pick it back up when the base is built, I understand that too. Decide with this information." Giving them a way out isn't weakness: it's the sentence that saves the most renewals, because it removes the feeling of being trapped.
When to refund, when to pause
Refund (in full or in part) when the failure was yours in the approach: you sold without auditing the starting point, you promised appearances or deadlines that couldn't be promised, or the diagnosis reveals something you should have seen before signing. It's cheaper than a burned client telling the story, and it forces you to fix the sales process — what can and can't be promised should be sorted before the proposal, not after.
Pause when the diagnosis is time or authority and the client can't sustain more months of investment: leave the base built and documented, agree which signal reactivates the service and keep a minimal measurement going meanwhile, because what's been built has to be watched.
Don't refund when the committed work is done and reported and what hasn't arrived is the result you never promised. There the answer is the diagnosis and the adjusted plan, not the refund. If you give in on this, your service becomes "guaranteed results or it's free", which is exactly the model you can't sustain in a volatile channel.
Frequently asked questions
How long without improvement is "too long"?
It depends on the branch and the sector, but a reference: by 3 months you should see intermediate signals (sources corrected, first citations, improvements in brand prompts); by 6, movement in appearances in at least some of the prompts. Six months with absolutely nothing in any AI calls for the conversation in this article, not another routine report.
And if the problem is my execution, not the client?
Check it before the meeting: was everything committed done, on time and to standard? Were the source corrections actually applied? If you find gaps, the honest conversation includes admitting them — "this month we were slower than we should have been on X, it's now sorted" — and making it right. Lying to yourself in the diagnosis only delays losing the client.
How do I avoid reaching this conversation?
With the expectations set at the sale: starting point measured before signing, phased goals and work KPIs. In our study of 9,865 SMEs, 91% appeared in only 1 of the 4 AIs (full study) — starting from that low is normal, and saying so on day one turns month four into a plan review rather than a crisis.
Doesn't showing a flat chart make me look bad?
The opposite, if you pair it with diagnosis and a plan. What destroys trust isn't the bad figure: it's the client discovering it on their own, or smelling that the report is hiding it. The flat chart with a why and a plan is professionalism; the dressed-up chart is a ticking bomb.
To have this conversation you need weekly, comparable data from the 4 AIs, not impressions — the four-branch diagnosis comes out of that. If you still measure by hand or in fits and starts, in Surfeo for agencies you have the per-prompt, per-AI history for each client, and you can start by seeing the real picture of any of them with the free AI visibility test.
