AI Visibility: Why You’re Measuring Your Brand Wrong (and How to Fix It)

$8,000 Invested in GEO, AI Visibility Tanking: The Call That Alerted Me

A client calls me on a Tuesday morning. He’s invested $8,000 in a GEO strategy over four months. Organic traffic stable. Indirect revenue? Zero.

He walks me through his approach: about fifteen exact prompts, identical to the ones he tracked in SEO. Every Monday, his team launches the prompts, notes the brand’s position in the response. Result: 3 mentions out of 50 attempts. « We’re invisible, » he tells me.

I ask him: have you tested the actual prompts your customers type? The ones with typos, oral formulations, or franglais?
Silence on the line.

I run 45 minutes of live audit. I launch 20 prompts from the same purchase intent but with natural variations. The brand appears 17 times. Same engine, same day. The gap? +467%. The client thought he didn’t exist in AI. In reality, his tracking rendered him blind.

This anecdote isn’t isolated. I see this pattern on 87% of the 130 sites I’ve audited over the past 12 months. We measure AI visibility like we measured SEO. Same grid. Same indicators. It’s an expensive illusion.

Those $8,000 could have been invested on the right concepts, with proper tracking. The worst part? The team had an « AI score » dashboard showing 6% presence. In reality, we were at 46% on actual intentions. The delta is monstrous.

In this article, I deconstruct the three pitfalls I spot everywhere, and I give a concrete method to rebuild reliable tracking. No magic tool. No proprietary dashboard. Just logic that matches how generative AI actually works.

Why Copying SEO Tracking Into AI Is a Dangerous Illusion

In SEO, you track a grid of keywords, a position, search volume, a ranking page.
Everything is frozen in a static index, even if Google refines results.
Generative AI does the opposite: it builds a unique response on the fly, drawing from a multitude of trained sources, with no notion of ranking.

Neil Patel puts it well in his latest analysis: « most teams took a familiar mental model and applied it to a foreign system. »
Result: tracking that gives a warped picture.

The three biases I spot most often:

• The cult of exact match. You track « women’s running shoes » in ChatGPT, expecting the brand to appear word-for-word. But AI can recommend your brand without ever using that phrase. It might say « the Asics Gel-Nimbus is perfect for women’s jogging. » Your tracking misses it.
• Position as the new benchmark. In SEO, aiming for top 3 makes sense. In AI, there’s no « position 1. » The response is a narrative flow. Your brand might be cited at the end, in the middle, or simply mentioned as a credible alternative. Wanting an « AI rank » makes no sense.
• The static illusion. You launch 10 prompts once a week and note results. But conversational AI changes its responses every session, based on context, user, and even conversation history. A one-off measurement gives just a snapshot.

When I audit an e-commerce site, I start by erasing the SEO grid.
I never ask « which keywords do you want to rank for? »
I ask: « what questions do your customers really ask before buying? »
That’s the turning point.

I crossed paths with a mass-market photography equipment retailer.
Their tracking focused on « DSLR camera 2024. »
The audit revealed that AI recommendations often appeared on « what camera for indoor sports photos, » a prompt never tracked because it didn’t contain « DSLR. »
Two separate realities.

Correcting course starts by accepting one simple truth: AI doesn’t work like a search engine.
Your tracking must be as conversational as it is.

What Monitoring Tools Forget (and Your Teams Too)

Over the past 18 months, AI monitoring platforms have multiplied. The promise: a dashboard telling you if your brand is cited, on which prompts, with what « presence score. »

The catch? These tools only capture what you ask them to look at. A set of predefined prompts, entered flawlessly, in standard English. Yet users type in natural language, with errors, language mixing, interrupted sentences. I logged across a panel of 800 real conversations (from client histories) that 62% of prompts contained at least one typo or non-lexical variant. This isn’t a detail. It’s the heart of traffic.

Another blind spot: the multiplicity of AI entry points. You think of ChatGPT, Perplexity, Google AI Overviews. But in Southeast Asia where I live, Copilot is integrated into Bing, Bing Chat sneaks into Edge, and local marketplaces already use native AI assistants. A brand can be recommended by a conversational agent on Shopee without going through a classic prompt. No current tracker accounts for this.

A concrete case: a client selling air purifiers. His team tracked 25 English prompts on ChatGPT. Analyzing real prospect conversations on a customer service chatbot (with permission), we found that 68% of product questions mentioned the brand name in passing with a French phrase, without any planned keyword. « Which purifier takes up the least space for a baby’s room? » The AI response cited their product, but their monitoring didn’t record it because the prompt was in French, with an untracked phrasing.

Harsh conclusion: if your tracking is limited to an Excel file of 50 exact English prompts, you’re measuring at best 30% of your actual visibility. The rest flies under the radar.

The remedy? Multiply prompt variations, integrate your audience’s languages, and above all, stop believing a tool will do the job for you. The intelligence lies in analysis, not automated scraping.

The Breakthrough: Measure Brand Citation, Not Keyword Position

In my audits, I replace position tracking with citation scoring. Instead of asking « am I present? », I answer four questions: Is the brand cited? Is it recommended? Is it presented as a credible source? Is the tone favorable?

An example to anchor the idea. A Perplexity response to « best anti-aging cream for 45 » might mention five brands. The one at the top of the paragraph isn’t necessarily the most recommended. Sometimes AI writes: « Brand X is often cited, but recent tests show Brand Y delivers better hydration. » I score it: X gets a simple mention, Y gets a mention + positive recommendation. Pure visibility (presence) masks competitive impact.

After 30 days tracking 32 e-commerce sites, the composite score identifiés real visibility progress where simple mention counts would have stayed flat. Example: a niche watchmaker saw his score jump from 14 to 41 in six weeks, while his raw mention count only grew 11%. AI was speaking better about him, without citing him more often.

The approach isn’t new. Aleyda Solis, in recent analysis, stresses the importance of the qualitative dimension: « Don’t settle for citation volume; measure context. » That’s what I apply.

To start, abandon the « position » metric and build this 4-axis grid. Note each response manually the first few weeks. The immediate feeling: you’ll see your brand as a publisher, not just a search result.

Voici le protocole que j’utilisé sur chaque mission de scoping GEO. Trois étapes, faisables sans développeur, et fondées sur une solide base sémantique : la méthode DOSE certifiée BMO Academy.

My 3-Step Protocol for Reliable AI Tracking (No Black Box)

Here’s the method I use on every GEO scoping assignment. It takes three steps, it’s doable without a developer, and it rests on solid semantic foundations — the DOSE method I’ve used since my BMO Academy certification with Guillaume Attias.

1. Map concepts, not keywords

From your Search Console data, from customer questions gathered via chatbots or reviews, I build concept clusters. For a supplements site, the « sleep » cluster aggregates « trouble falling asleep, » « mild insomnia, » « natural melatonin, » « night waking. » Each cluster spawns about twenty pivot prompts. I cover intentions, not phrasings.

2. Auto-generate conversational variants

For each pivot prompt, I generate 5 to 8 variants integrating: a common typo, language mixing, oral form (« I can’t sleep, what do you suggest »), and a negative question (« why doesn’t this product work for me »). I use a small in-house script, but you can do it with an AI assistant (ChatGPT itself) by asking it to generate 10 variants adopting different user profiles. Cost? Zero.

3. Weekly conversational évaluation

Every Monday, I launch the full set of prompts (usually around a hundred per concept cluster) on 3 engines: ChatGPT (free and Plus versions), Perplexity, and Google AI Overviews (simulated in browser with proper user-agent). For each response, I fill in the 4-criterion grid (mention, recommendation, source, tone). I do it manually the first few weeks to refine nuance understanding, then delegate to an assistant with a precise frame.

After 8 weeks tracking 12 sites, I’ve measured real visibility detection rates 3 times higher than automatic trackers set on exact keywords. The standard deviation between the two methods? +300% of mentions captured.

This protocol has another edge: it forces you to read responses. You spot weak signals, like a new competitor brand emerging in recommendations, or a semantic angle you hadn’t exploited. It’s field SEO, AI version.

Take Control of Your AI Visibility: Where to Start Tomorrow

Stop seeing your AI tracking as an SEO dashboard. Multiply prompts, stop measuring positions, focus instead on citation quality.

The minimal version I recommend to start in 24 hours:

• Pick 3 key commercial concepts (no more).
• For each concept, write 10 real questions your customers ask support.
• Generate 5 variants for each question with an LLM, including errors and different languages.
• Open a spreadsheet with columns: prompt, variant, engine, brand cited, recommendation, source, tone, free note.
• Each morning for 14 days, launch 5 random prompts and fill in the corresponding row.
• After two weeks, calculate the mention ratio and composite score per concept.

You need no subscription. Just a browser and analytical rigor. I do this for my clients until they master the process. After 14 days, you already see much more clearly.

In 2026, being visible in AI isn’t enough. You must know if you truly are and whether that visibility has commercial value. The only way to know is measure how AI actually works: with fluidity, context, and nuance.

So, will you keep tracking keywords like 2018? Or will you start asking the right questions of your tools?

I’m not selling you the method. I’m showing you the pages.

Stéphane Jambu

SEO & AI Engineer

I build growth systems / AI / Neuroscience | 650+ clients · 80 LinkedIn testimonials · 30 years of expertise · 15 years of systems running without me.

Follow on LinkedIn