Two families of metrics to measure the impact of LLMs
Measuring GEO (Generative Engine Optimization, that is, search visibility in generative engines) comes down to tracking your brand's place in the answers of ChatGPT, Perplexity, Gemini, Copilot or Google's AI Overviews. The whole framework rests on one simple distinction. Presence metrics measure that place in the answers: they react fast and steer the day-to-day work. Outcome metrics measure the real effect on revenue: leads, sign-ups, sales, with weeks of lag. The former justify the effort, the latter justify the budget. You need both.
Presence metrics: prompt tracking
Prompt tracking consists in running, every day and across several models, a library of questions representative of your business and product goals, then recording what the answers say about your brand. If you sell e-invoicing software, for example, these are the questions a CFO asks before choosing: "which e-invoicing software should an SMB choose?", "which solutions comply with the new mandate?". Draw them from sales meetings, support tickets and Search Console queries rephrased in natural language. Keep the prompts that name your brand separate: they almost always mention you and inflate every score.
A prompt tracker such as Peec, Profound or Otterly automates this collection and records three data points in each answer: mentions (your brand is named), rank (your brand's place among those the answer cites) and citations (the pages the model displays as sources, whether yours or a third party's).
From mentions and their rank I derive five metrics. Citations come in further down, as a qualitative read.
| Metric | How it's computed | What it tells you |
|---|---|---|
| Share of voice | Your mentions over those of all tracked brands, yours included | How you perform against others. Means nothing in business terms by itself, but its trend validates your efforts, or not. |
| Visibility | Share of the prompts run in which you are mentioned | How well you cover the scope of your prompt set. A good steering metric. |
| Position | Average rank of your mention among the brands cited, counting only the prompts where you appear | Being quoted first probably means you rank higher in credibility. Read on its own, it flatters: disappearing from the prompts where you ranked poorly improves it. |
| Weighted mention | Average of the inverse rank across all prompts, absences included, on a scale of 100 (formula below the table) | Combines coverage and rank in a single number. Tracked daily, it is the steering metric, the most telling of the lot. |
| Weighted share of voice | Your weighted mention over the sum of the weighted mentions of all tracked brands | Your weight in the answers against the market, coverage and rank combined. Richer than classic share of voice, and complementary. |
The weighted mention deserves its formula, since it is the one steering the work. For each prompt, your mention is worth the inverse of its rank: 1 if you are quoted first, 0.5 second, 0.33 third, 0 if you are absent. The average of these scores across the whole library, multiplied by 100, gives the metric. Example over 100 prompts: quoted first 30 times, second 20 times, third 10 times and absent from the other 40, that is 30 × 1 + 20 × 0.5 + 10 × 0.33 + 40 × 0 = 43.3 points over 100 prompts, hence a weighted mention of 43 out of 100. The 40 absences weigh in: that is what sets this metric apart from position. Compute it model by model before averaging: ChatGPT and Perplexity do not rank you the same way, and the gap between the two often teaches you something.
Add two qualitative reads. Citations first: a cited page influences the model's narrative. Spot the ones that work and use them to justify your investments or to prioritise your actions: which discussion threads to influence first, which video creators to contact. Watch your competitors' citations too. Sentiment next: how the models talk about you. If you regularly brief third parties who will talk about your brand, track the effect of each event, like a newly earned label, on overall sentiment or split by product.
Four complementary sources: search engines, logs, panels, AI Overviews
These four sources measure neither presence nor outcome: they shed light on the mechanics in between, what the bots read, what the engines display, where visitors come from. Check them once a month, or when a prompt tracking number moves for no visible reason.
| Source | What it gives you |
|---|---|
| Search engine data | Bing Webmaster Tools and Google Search Console each publish data on the place of your pages in their generated answers, Copilot on one side, AI Overviews and AI Mode on the other. Mostly useful to justify budget allocation, or to dig into something specific. |
| Server logs | Filter the LLM bots in your logs and separate two families: training and indexing bots (GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot), which feed the models and their indexes, and answer bots (ChatGPT-User, Perplexity-User, Claude-User), which read a page at the moment a model composes an answer. The latter prove, page by page, that you serve as a source. And when you ship a technical improvement to help models understand the site, this is where its effect on the crawl side shows. |
| Third-party panels | Some panels, like Similarweb, estimate what share of your visitors also used ChatGPT or Perplexity the same day, and how many visits each assistant sends you compared with your competitors. Estimates to take with caution, but the angle, external and comparative, is valuable. |
| AI Overview tracking | Your rank tracker (Semrush, Ahrefs or SE Ranking offer it) shows, keyword by keyword, whether an AI Overview appears and whether your site is cited in it. A good steering metric to measure the effect of a change made to the site: new content, new formatting. |
Outcome metrics: the real impact on your leads
Declarative data first. The question "how did you hear about us?" measures the real business impact, and its verdict often diverges from your attribution model's. Without it you miss the most common scenario: someone reads a recommendation in an LLM, types your brand into Google, clicks a paid or organic link. No attribution model sees that chain.
Analytics next. In GA4, create a custom "LLM" channel group defined by a regular expression on the session source: chatgpt.com, perplexity.ai, copilot.microsoft.com, gemini.google.com, claude.ai, chat.mistral.ai. Without it, these visits drown in the "Referral" channel. When a visitor arrives by clicking a link in an answer, you can attribute it. That number is a floor: clicks from AI Overviews and AI Mode arrive as Google organic traffic, indistinguishable from classic SEO, and the assistants' mobile apps often send no referrer, so those visits fall into direct traffic. Complete the picture with cookies: if a prospect visits your site, goes through an LLM, then comes back, the Conversion paths report shows that step as an intermediate touchpoint, whatever attribution model you have chosen.
Indirect signals last. An unexplained uplift in brand and direct traffic often reflects the same scenario as the declarative data: a recommendation read, then a search. Run a daily analysis of your sales team's call transcripts to catch prospects who mention an LLM as a source; a small automation is enough, a script that asks a model a single question about each transcript, "does the prospect say they discovered the brand through an AI?", and ticks a CRM field when the answer is yes. And keep an eye on declarative market studies, public or paid, that look into how people search: always insightful.
| Signal | What it captures | Its limit |
|---|---|---|
| Declarative | The real channel through which prospects found you | Relies on the respondent's memory and honesty |
| Analytics (GA4) | Direct clicks from LLM answers | Blind to a recommendation followed by a brand search, and to clicks without a referrer |
| Indirect signals | Brand traffic, sales transcripts, studies | Correlations, never proof on their own |
Bringing it all together, and where to start
The last link is the simplest, provided you follow an order. On day one, set up what is free and whose history cannot be recovered: the "how did you hear about us?" question in your forms and the LLM channel group in GA4. In the first week, build the prompt library and start tracking. In the first month, plug in the complementary sources.
Then gather all this data in one place, a spreadsheet is enough at first, and hand an AI assistant, Claude in my case, a weekly analysis, with a short trend recap delivered as a Slack message. Three reading rhythms: the weighted mention every day, by the person doing the work; the recap every week, by the marketing team; presence and outcome side by side every quarter, in front of management, because that comparison is what decides the budget. Measurement only matters if someone reads it every week.
In 2026, the impact of LLMs on lead generation can't be read off any ready-made dashboard. You rebuild it metric by metric, presence on one side, outcome on the other, until both curves tell the same story: the weighted mention rises one quarter, the share of leads who name an AI rises the next. As long as they diverge, you know where to look: presence without outcome, you are cited on prompts that do not bring in your buyers; outcome without presence, your library does not cover what they really ask.
Frequently asked questions
More analyses on GEO measurement and generative search are available in the journal's articles.