How to get recommended by ChatGPT and Perplexity: co-mentions and what drives AI recommendations

An AI engine can describe your product accurately and still never put it on a shortlist. That gap is the thing most content teams have not measured, and a study published June 11, 2026 on Search Engine Land finally puts a number on it. Across 12 athleisure brands tested over 14,140 API runs, Nike showed up in 71% of "best athleisure" recommendation prompts while New Balance and Reebok appeared in 0%, despite all three carrying the identical Knowledge Graph description, "Footwear company" (per Search Engine Land). Same entity definition, opposite outcome. The difference was co-mention density: who each brand appears alongside in third-party editorial content.

This matters more now because the click is disappearing. Google searches in the US ended without a click 68.01% of the time across the first four months of 2026, up from 60.45% in 2024, per SparkToro clickstream research reported by Search Engine Land. When two-thirds of searches never leave the results page, the AI answer is the result. Getting cited inside that answer is the new ranking. Below: the mechanism behind co-mentions, five tactics that move your position, and a 30-day audit you can run with a crawler like Firecrawl (/go/firecrawl) to track whether it is working.

What a co-mention is, and why recommendation engines weight it

A co-mention is two brands appearing in the same piece of third-party content. Not linked to each other. Just named together, in a roundup, a comparison, an analyst report, a retailer's category page. The study mapped these pairs across UK-indexed editorial sources and found tight clusters: lululemon and Alo Yoga co-occurred 534 times, lululemon and Nike 482 times, Alo Yoga and Nike 449 times (per Search Engine Land). Those brands keep landing in the same articles, so the model learns to treat them as one category. New Balance sits outside that cluster. Its co-mention density lives in running and performance content, not athleisure, so the engine never built the bridge.

The mechanism is pattern-matching, not reasoning. An LLM answering "best athleisure brands" does not evaluate each brand on the merits and decide who belongs. It retrieves the brands that consistently appear together in athleisure content and returns those. If your brand was never in that company, the association was never formed, and you do not surface. The study's authors frame it through Jason Barnard's point: if A plus B should equal a category, you have to construct that path in external content explicitly, because the model will not infer it for you.

This is why recognition and recommendation pull from different sources. When a user types your brand name, the engine leans on your own pages. When a user asks for a recommendation in a category without naming anyone, the engine leans almost entirely on third parties.

Prompt type	What the user asked	Own-brand citation rate	Third-party citation rate	Implication
Recognition ("What is [Brand]?")	Named your brand	ChatGPT cited own-brand content 49% (Perplexity 36%, Claude 23%)	The rest	Your About page, schema, and category pages do the work. This you can fix on your own site.
Recommendation ("Best [category]?")	Named no brand	Drops to 18% on ChatGPT, effectively zero on Gemini, Claude, Perplexity, and Google AI Overviews	82% to 100% across all five systems	You cannot fix this on your own site. It is a category-positioning and PR problem.

Citation source by prompt type, from the 12-brand athleisure study, Search Engine Land, 2026-06-11.

Read the bottom row again. For decision-stage queries, between 82% and 100% of what these five engines cited came from sources you do not own (per Search Engine Land). Tuning your homepage copy will not move a number that homepage copy does not feed. The same week, Search Engine Land argued that AI-written content reads generic and gets discounted precisely because it lacks the firsthand specificity engines reward (per Search Engine Land), which is another reason owned content alone underperforms at the recommendation stage. The work that moves you happens off your domain.

One caution before the tactics: this is a single study, one category, 12 brands, UK geography. Treat the specific percentages as evidence from that sample, not a law of physics. The generalizable claim is the structural one. Co-mention density in the right category cluster correlates with recommendation visibility, and the study does not publish a magic threshold for how many co-mentions you need. Neither will this guide.

Five tactics that shift your co-mention position

These map directly to the content types the study flagged as high-signal. None of them require a hyperlink back to you. An unlinked mention in the right editorial context still builds the association.

Get into roundup comparisons that name your category leaders. Being in a "best of" list alongside the brands that define your category does more for your concept-graph placement than a standalone profile, because the cluster signal comes from sharing an article with those brands.
Land in analyst and sector reports that group tools by category. Category-level reports that list peers together are dense co-mention sources, and they place you in the cluster in a way solo coverage does not.
Do podcasts where the host frames you against named competitors. A bio that says you "compete with lululemon and Gymshark in premium athleisure" indexes the co-occurrence; "a performance apparel company" does not.
Fix your retailer and directory categorization. A major retailer's category page is external content, so being stocked and grouped alongside category leaders is itself a co-mention signal the engine can retrieve.
Build editorial relationships that produce "X vs Y" content. Direct comparison pieces pair your name with a leader's by design, which is the strongest, most explicit version of the co-occurrence signal.

The throughline: aim for visibility in the right company, not just visibility. A press hit calling you "performance apparel" in isolation does little; the same hit listing you next to the three brands that own your category does the structural work.

The 30-day co-mention audit

You cannot fix what you have not measured, and the measurement here is unusual: you are auditing third-party content, not your own. Run this over a month.

Days 1 to 5: capture your baseline citations. Run recommendation prompts for your category on ChatGPT, Gemini, and Perplexity. Use phrasing a buyer would use, not your brand name. "Best AI coding tools for enterprise teams." "Top customer-support automation platforms 2026." Log every source the engine cites and whether your brand appears at all.

Days 6 to 12: map who you are cited alongside. Open each cited source and record which competitors appear in the same piece. This is your current cluster. If the brands that dominate your category are present and you are not, you have found the gap. If you appear but in a different cluster (the way New Balance sits in "running," not "athleisure"), you have found a positioning problem.

Days 13 to 20: build a target list. Identify the 10 to 15 editorial properties that own your category's recommendation queries, the roundups, comparison sites, analyst blogs, and retailer category pages the engines actually cite. These are your pitch targets. The goal is to appear in those specific pieces, named alongside your category leaders.

Days 21 to 30: stand up a monthly tracker. Co-mention position drifts as editorial content updates, so measure it on a schedule. Write a script that fetches each of your 15 target pages and counts how many times your brand appears in the same page as each named competitor. A basic fetch call breaks on JavaScript-rendered editorial pages and infinite-scroll roundups, which is most modern publisher templates. To pull clean, rendered text from those pages on a monthly cron, Firecrawl's scraping API (/go/firecrawl) returns the page as markdown you can run a simple co-occurrence count against. There is no GEO dashboard here and you do not need one. You need rendered text, a competitor name list, and a counter that you diff month over month.

# Monthly co-mention tracker: count how often your brand appears
# on a target page alongside each named competitor.
import os, requests

FIRECRAWL_KEY = os.environ["FIRECRAWL_API_KEY"]
BRAND = "yourbrand"
COMPETITORS = ["competitor-a", "competitor-b", "competitor-c"]
TARGETS = [
    "https://example-roundup.com/best-tools-2026",
    # ... your 15 target editorial pages
]

def fetch_markdown(url: str) -> str:
    r = requests.post(
        "https://api.firecrawl.dev/v1/scrape",
        headers={"Authorization": f"Bearer {FIRECRAWL_KEY}"},
        json={"url": url, "formats": ["markdown"]},
        timeout=60,
    )
    r.raise_for_status()
    return r.json()["data"]["markdown"].lower()

for url in TARGETS:
    text = fetch_markdown(url)
    if BRAND not in text:
        print(f"MISSING  {url}  (brand not on page)")
        continue
    paired = [c for c in COMPETITORS if c in text]
    print(f"PRESENT  {url}  co-mentioned with: {paired or 'none'}")

Run this on the first of every month and track two numbers per page: are you present, and which competitors you share the page with. A page where you are present but co-mentioned with "none" of your category leaders is the New Balance problem in miniature.

Schema and FAQPage markup: a supporting signal, not the lever

Structured data does not buy you a recommendation, but it helps the recognition half of the equation, where your own pages carry 23% to 49% of citations depending on the engine (per Search Engine Land). Clean schema makes your entity legible so that when an engine retrieves your page, it can parse what you are. FAQPage markup specifically gives answer engines pre-chunked question-and-answer pairs they can lift directly, which is the format AI Overviews tend to reward.

Add it to a page that already answers real category questions in plain prose. The markup describes content that exists; it does not substitute for it. The full set of on-page moves that earn recognition-stage citations, statistics, extractable answers, clean entity data, sits in our companion 2026 GEO playbook; this guide is about the off-site half that playbook does not cover.

{
  "@context": "https://schema.org",
  "@type": "FAQPage",
  "mainEntity": [
    {
      "@type": "Question",
      "name": "What does YourTool do?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "YourTool is a [category] platform for [audience] that [specific outcome]."
      }
    },
    {
      "@type": "Question",
      "name": "How does YourTool compare to [Category Leader]?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "YourTool focuses on [differentiator], while [Category Leader] focuses on [their strength]. For [use case], YourTool fits because [reason]."
      }
    }
  ]
}

Note the second question. Naming a category leader inside your own structured data will not on its own build the third-party co-mention signal that recommendation queries run on. But it does keep your owned content honest about the category you are competing in, which is the same discipline the off-site work demands.

What to do this week

Run three recommendation prompts for your category on ChatGPT, Gemini, and Perplexity, without naming your brand, and log every source cited. Open those sources and write down which competitors appear in each. If your category's leaders are in the room and you are not, you have your answer: the work is off your domain, in the editorial content that defines your cluster. Pick the five highest-signal properties from that list and start there. The recognition fixes (schema, FAQ markup, clean entity data) are worth doing, but they are not what gets you recommended. Co-mention density is.