Why mention rate became the vanity metric of AI search, and the one column that fixes it

In 1985, a PR firm would mail you a bound book of press clippings and bill you by the thickness. Nobody read the clippings. The book proved you existed in print, which felt like proof of something bigger, and the invoice went out either way. Mention rate is that book with a dashboard bolted onto it.

First, what mention rate actually measures

Here is the plain definition. You pick a list of prompts your buyers would plausibly type into ChatGPT, Gemini, Perplexity, or Google AI Mode. Something like “best CRM for a 20 person sales team.” A tracking tool runs those prompts on a schedule and records whether your brand name appears anywhere in the generated answer. Mention rate is the percentage of runs where your name showed up.

That is the entire measurement. It is a string match. It does not read the sentence your name is sitting in, and it does not know whether the model was recommending you, listing you as an also-ran, or holding you up as the thing the buyer should steer clear of.

The four ways your brand can appear in an AI answer

Recommended. The model names you as the pick, or as one of two or three picks, and gives a reason. This is the only version that behaves like a referral.

Mentioned neutrally. You appear on a list of vendors in the space. No verdict attached. You are furniture.

Named as the thing to avoid. You appear as the expensive one, the one with support problems, the one everybody migrated off of, or the cautionary example the model includes to make its actual recommendation look balanced.

Cited. A mention is your name in the text. A citation is a clickable link back to a source. They are tracked separately because they do different jobs, and a brand can easily get one without the other.

Every one of those four counts as a “mention.” Three of them are not wins. One of them is actively costing you money.

The sentiment split almost nobody puts on the dashboard

Geostar’s February 2026 sentiment analysis of brand mentions across AI responses found the breakdown was 80.6% neutral, 18.4% positive, and 1% negative. Read that again. Roughly four out of five times your name appears, the model is reading it off a list.

The same analysis found that only 28% of AI answers give a brand both a mention and a citation. So the majority of your “visibility” is a name drop with nothing behind it, and no path for the reader to go verify anything or click through to you.

If your mention rate went from 12% to 19% last quarter and you have no idea how that split across the four categories above, you do not actually know whether the quarter was good.

Where these answers are being assembled from

This is the part that should make you uncomfortable, because the sources feeding AI answers are heavily weighted toward places where unhappy people go to type.

A Semrush analysis of 150,000 citations across 5,000 randomly selected keywords found Reddit referenced in 40.1% of AI answers, ahead of Wikipedia at 26.3%, YouTube at 23.5%, Yelp at 21%, and Facebook at 20%. A separate 5W Citation Source Audit for Q1 2026, drawing on Similarweb data covering roughly 600,000 citation events, found Wikipedia (13.15%) and Reddit (11.97%) together supplying more than a quarter of all ChatGPT citations in the United States. The Wall Street Journal, the New York Times, and Bloomberg did not appear in the top 20.

Tinuiti’s Q1 2026 AI Citations Trends Report tracked nine commercial categories across seven AI platforms and found Reddit citation share up at least 73% between October 2025 and January 2026, reaching 24% of all Perplexity citations and 44% of the social citations inside Google AI Overviews.

An EMGI study of 1,486 B2B software buying queries across 18 categories found Reddit present on Google’s AI-enhanced results page 81.6% of the time. On bottom-of-funnel queries, the ones where somebody is about to pull out a credit card, that figure hit 94.1%.

Nobody has ever posted a thread titled “AMA: My Renewal Went Smoothly.” The threads that get written, upvoted, and then scraped are complaint threads, switching guides, and “we moved off X, here is why” posts. Those are the pages driving your mention count up.

Your buyers are hunting for the bad news on purpose

This is not a hypothetical risk. It is documented shopping behavior.

82% of shoppers deliberately seek out negative reviews to establish whether a company is credible, because a wall of five star ratings reads as filtered (Capital One Shopping, 2026). 66.2% say negative reviews directly change their purchase decisions (PissedConsumer, 2026). BrightLocal’s 2026 survey found 95% of consumers read reviews before buying and 77% say negatives make them less likely to use a business. Northwestern’s Spiegel Research Center has repeatedly found purchase likelihood peaks in the 4.2 to 4.5 star range rather than at a perfect 5.0.

An AI model summarizing your category is doing exactly what that buyer would do, only faster and with no patience for nuance. It reads the complaint thread and compresses it into one clause about you.

Why the stakes are higher than they were with blue links

In a traditional search result you got ten listings and the user picked. In an AI answer there is no position four. There are two or three names and a verdict.

Search Engine Land’s 2026 zero-click study found roughly 83% of queries showing an AI Overview end without a click, rising to about 93% inside AI Mode. Seer Interactive’s analysis of 25.1 million impressions across 42 organizations found brands cited in AI Overviews earned 35% more organic clicks and 91% more paid clicks than uncited brands on the same results page. The Opollo 2026 AI Search Benchmark, covering 312 B2B technology firms, reported AI-referred traffic converting at 14.2% against 2.8% for Google organic.

So the traffic that does come through is worth roughly five times as much per visitor. Which means the sentence attached to your name is worth roughly five times as much too.

The fix: add one column to the log you already keep

You are already logging mentions. Add a sentiment column and score every single one three ways: recommended, mentioned neutrally, or named as the thing to avoid. Then track those three lines separately over time instead of summing them into one number.

Practical build, one afternoon of work:

1. Pull 50 to 200 prompts per platform. Anything fewer and the sample is noise. Weight them toward the queries tied to demos, trials, and pricing.

2. Capture the full sentence, not the name. Your log needs the clause your brand sits inside. “X is solid but pricey for small teams” is a different asset than “X is the one to beat.”

3. Score it, then log the source. Every negative score gets a link to the page that caused it. That link list is your actual work queue.

4. Report three lines on the chart. Recommended, neutral, avoid. Never a single stacked total.

Publish the ugly number anyway

Here is the finding you should expect, and the reason most teams quietly skip this exercise. A meaningful share of companies celebrating rising AI visibility will discover their negative mentions are the fastest growing line on the chart, because the third-party pages responsible for the increase are complaint threads and competitor switching guides.

Publish it regardless. A negative mention traced to a specific Reddit thread or a specific G2 review pattern is a fixable problem with an address. A rising blended mention rate is a feeling.

And treat all of these numbers as perishable. PromptWatch tracked Reddit’s share of ChatGPT citations falling from a 3.83% average down to 0.52% in August 2026, an 86.4% relative drop inside two weeks, after OpenAI changed how the model queries the web. Semrush observed a similar collapse in September 2025, from roughly 60% of prompt responses to about 10% over a 13-week, 230,000-prompt tracking study. Build the ledger so it survives the platform moving underneath it.

Being mentioned is a fact about a string of text. Being recommended is a fact about your business. Only one of them shows up in the pipeline.

Sources

Geostar AI brand mention sentiment analysis (February 2026) · Semrush / Statista, 150,000 citations across 5,000 keywords · 5W Citation Source Audit Q1 2026 (Similarweb, ~600,000 citation events) · Tinuiti Q1 2026 AI Citations Trends Report · EMGI Group, 1,486 B2B SaaS buying queries · PromptWatch citation monitoring, August 2026 · Semrush 13-week, 230,000-prompt tracking study · BrightLocal 2026 Local Consumer Review Survey · Capital One Shopping online review research, 2026 · PissedConsumer 2026 review trends report · Northwestern Spiegel Research Center · Search Engine Land 2026 zero-click study · Seer Interactive, 25.1M impressions across 42 organizations · Opollo 2026 AI Search Benchmark, 312 B2B firms.