This July, Semrush asked 2,338 US adults how AI chatbots shape what they buy. Among those who use AI, 57.5% had bought a product because a chatbot recommended it. The same share had decided not to buy something because of what a chatbot told them.

AI talks as many people out of purchases as into them. A Google search used to provide the wine list. Now the AI chatbot is the sommelier. It suggests the bottle, and it tells you which one to skip.

The industry’s response to zero-click search has been to reconstruct traditional SEO KPIs in the form of mentions and citations. Mentions count how often AI names your brand – a stand-in for impressions. Citations highlight how often the answer links to your site, a stand-in for clicks. Neither tells you whether ChatGPT told a buyer to pick you or avoid you, or if anyone clicked the link.

At Aftergrowth, we measure recommendation rate instead: how often an AI answer tells a buyer to choose you.

Mentions are the new impressions

SEO was judged by the clicks it drove. Its job was to send traffic to a destination, and clicks and CTR were reasonable primary KPIs because they led to conversions, purchases and revenue. They looked great in default last-touch attribution and held up fairly well in most MMM.

In the mad dash to zero-click, everyone agreed SEO needed new metrics. The industry picked the easiest (laziest) substitutes it could find. Mentions became the new impressions. Citations were supposed to become the new clicks, except hardly anyone clicks them. A 2025 Pew study found users clicked a link inside an AI summary on just 1% of visits.

And yet, most brands and agencies are chasing some combination of mentions, citations, and overall visibility. Mentions earn their place at the top of the funnel. The mistake is treating them as the finish line. A mention can be negative, or one name in a list of ten, and it still counts the same.

In paid media, growth marketers spent the last decade moving to KPIs that predict incremental revenue. Put impressions at the top of a paid media report today and a good growth marketer will channel Shania Twain: “That don’t impress me much.” Too many of those same marketers haven’t applied that lens to GEO.

Recommendations are the best available proxy for conversions, purchases and revenue, and the strongest evidence comes from Scrunch AI. Its researchers linked an opt-in clickstream panel to the same users’ ChatGPT, Claude and Gemini conversations.

In other words, recommendations drive brand search and direct site visits, both of which have a throughline to revenue.

We expect future data to show those numbers understate the effect. Meanwhile, conversations with friends, colleagues, and even older family members keep circling back to a grill they bought, a restaurant they tried, a book they read, or software they shortlisted because an AI answer recommended it and showed its work.

The recommendation ladder

Recommendations come in degrees. From best to worst, here is where an answer can put you, and what each metric actually counts.

The bottom rung is the one most dashboards miss. A brand that gets recommended against still logs a mention, and it’s worse off than a brand that never came up.

We’ll cover the methodology, including the edge cases, in a separate post.

The case for absolute recommendations is less settled. The Scrunch study found the lift was flat regardless of where the brand appeared in the answer. Our hypothesis is that position matters in proportion to how many options a buyer will actually evaluate. Fourth on the list for toothpaste probably does very little. One of four for B2B software is a seat at the table. In fact, G2 reported that 80% of software buyers were shown to have purchased from their initial AI-delivered shortlist.

The case against recommendation rate, and why it still wins

Recommendation rate faces three fair objections. The first is effort. Few AI visibility tools report it out of the box, so it takes custom setup. That’s a cost, but it’s also an edge. Competitors who stop at mentions are optimizing toward a number further from revenue.

The second is noise. Sampled prompts aren’t real demand, and the same prompt can return a different answer an hour later. Both are true, and any GEO report should be read as directional. But the objection applies just as much to mentions, citations and every visibility score on the market. The fix is to run prompts repeatedly, track the trend and anchor on performance relative to competitors. If a competitor is recommended twice as often as you across hundreds of runs, that gap is real even when any single answer isn’t.

The third is the sharpest. Most AI visibility tools already track sentiment, so isn’t a recommendation just a mention with positive sentiment? Most platforms report sentiment in aggregate, as a summary of how a brand is described across a prompt set. An answer can describe you warmly and still send the buyer somewhere else. In our test below, ChatGPT recommended monday.com for “teams wanting highly visual workflows” in the same answer where it told a typical SMB to start with Asana.

Shifting the focus of GEO

As covered in The Controllables, choosing the primary objective is one of the most consequential decisions in any marketing effort, because what you measure is what you’ll optimize toward. Optimize toward mentions and you’ll grow the top of the funnel with no guarantee any of it reaches a shortlist. Optimize toward citations and you’re chasing links almost nobody clicks. Both still earn a place in the report. Mentions show how far your reach extends, and citations show which sources shaped the answer. Recommendation rate tells you whether you’re winning.

Measured this way, a GEO program comes down to three questions, asked prompt by prompt and segment by segment:

1. Where are you losing? Which topics, funnel stages, customer segments and AI platforms leave you out of the answer, or argue against you?

2. Should you be winning there? Check every gap against product fit. If your project management software isn’t built for 1,000-person enterprises, missing those prompts isn’t a problem. If it’s a great fit for 100-person startups and you’re still absent, it is.

3. Why are you losing? Look at who gets recommended instead and which sources shaped that answer. This is where citations earn their keep, as evidence of what the model was reading.

Most GEO roadmaps skip the second question. It’s the one that keeps you from chasing prompts you shouldn’t win.

Recommendation rate in action

To make this concrete, put yourself in the seat of a GEO manager for monday.com, a category leader in project management software. We ran four nearly identical prompts through logged-out ChatGPT, asking for the best project management and collaboration software for a different kind of company each time. Four prompts make a demonstration. A real program would run hundreds.

Monday.com was recommended in two of four answers, first in neither, and was one of four to six brands each time. Asana, its biggest competitor, has a 100% recommendation rate and was the outright pick in two.

The pattern matters more than the totals. Monday made the list on both generic prompts. Every time the prompt added context, a specialist took its seat: Linear and Jira for the AI startup, Teamwork.com for the web design firm. Asana survived all four because ChatGPT files it as the tool for cross-functional teams, a safe answer for almost any company. Monday gets described as “highly configurable” and “very visual.” Those are real strengths, but they don’t map to a specific buyer, so Monday never wins a specific prompt.

Then check whether the pattern holds on other platforms. It rarely does. On the same four prompts, Gemini never named an outright pick for any brand. Monday was recommended in three of four answers, always as part of a longer list, and Asana was far less dominant. Monday has a different problem on each platform.

Should Monday win there?

The AI-startup prompt may simply be Linear’s to lose. An engineering-heavy team picking Linear or Jira is a product-fit outcome, and chasing it could be wasted effort. The web design firm is different. When we followed up with “what about Monday?”, ChatGPT reversed course and said Monday “could be a very strong fit,” listing workload management, time tracking and budget dashboards. Take the reversal with a grain of salt, since models agree with almost anything you suggest. But Monday’s own team knows whether it serves agencies well, and if it does, this is the gap worth closing.

Why Monday loses

Every lost prompt has a cause: how you describe the product on your own site, how review sites rate you, or a shortage of third-party content (reviews, comparisons, YouTube walkthroughs) showing you working for that kind of customer. The sources and the model itself will tell you which.

Start with the sources. For the under-100-person prompt, ChatGPT built its answer from three small comparison sites: Dupple, The Productive Founder and Resources Rework. That’s good news for Monday. Getting into a listicle’s next update is far easier than moving a G2 rating, and so is reaching the reviewers and creators those sites draw on.

Then ask what the model didn’t read. On the AI-startup prompt, when we asked whether it had evaluated review sites, ChatGPT said it did “a fresh check this time” and that the review-site evidence “changes my earlier framing somewhat.” That suggests the review sites B2B buyers trust most played little part in the first answer.

Finally, ask the model why. Asked why it initially left Monday off the web design shortlist, ChatGPT said it had weighted “purpose-built for an agency” over “flexible general-purpose work management,” and named client-project financial management as Monday’s biggest gap. Its own description of Monday includes time tracking and budgets, which suggests the gap is at least as much in how Monday’s agency story shows up in the sources models read as in the product. Agency-specific content, case studies and review coverage on budgets, billable time and profitability are where Monday’s roadmap would start. Peec.ai has a new objections feature that will highlight the reasons why buyers avoid your product.

Measure the verdict

The sommelier-ization of search is real: credible, authentic, confident recommendations drive sales. In the wine-infused classic Sideways, Paul Giamatti’s Pinot Noir-loving character utters the famous line, “I am NOT drinking any f’ing Merlot!” In the subsequent years, Pinot sales went up 16% while Merlot showed slight drops.

AI is now pouring for every buyer in your category, and it talks as many of them out of purchases as into them. Your GEO reporting should tell you which side of that you’re on. Start by benchmarking recommendation rate against your top competitors on the prompts that matter, then compare it with your mention and citation numbers. The gap between them is your roadmap.

Mentions get you on the wine list. Recommendations get your bottle poured.