Mentions, Citations and Recommendations

A mention is your brand named in an AI answer. A recommendation is the answer suggesting you as a good choice for the asker's situation. A citation is a linked source in the answer. They're three different outcomes, they move for different reasons, and blending them into one "visibility" number hides the information you need. Lesson 3 of GEO 101 gives you scoring rules you can apply consistently, ten edge cases, a scoring sheet you'll use for the rest of the course, and MentionX data on how often these outcomes come apart.

In this lesson

• Why one visibility number isn't enough
• Mentions: the definition and the rules
• Recommendations: the hardest one to score
• Citations: yours, about you, and everyone else's
• Ten edge cases
• The scoring sheet
• From MentionX data: how the outcomes come apart
• Worked example: scoring three Ledgerline answers
• Which outcome to prioritise
• Exercise: score ten answers
• Quiz
• Sources

Anything marked "Our view" is our opinion from scoring AI answers, not documented engine behaviour. Ledgerline, the company in the worked example, is fictional.

Why one visibility number isn't enough

Imagine two brands, each "visible in 40% of answers".

Brand A is named in 40% of answers, almost always in a list of eight vendors. It's never singled out and never linked.

Brand B is named in 40% of answers too. In half of them the assistant says "for your situation, start with B", and B's own comparison page is the linked source.

Same headline number, very different businesses. Brand A has an awareness problem disguised as success. Brand B is winning shortlists. The only way to see the difference is to score the three outcomes separately. That's what this lesson sets up.

Mentions: the definition and the rules

A mention is your brand or product name appearing in the answer text. Rules:

1. Count once per answer. Being named three times in one answer is still one mention. Note repetition in the notes column if it matters.
2. Count negative and neutral mentions. "Ledgerline has had complaints about support" is a mention. Sentiment is a separate column.
3. Record position. If the answer has a list, note where you first appear (1st, 2nd...). If it's prose, note whether you're named in the first paragraph.
4. Product names count for the parent brand if your buyers would recognise them. Decide this once and write it down.
5. Citations alone don't count as mentions. If your URL is in the sources but your name isn't in the text, that's a citation without a mention, which is a separate edge case (below).

Recommendations: the hardest one to score

A recommendation is the answer suggesting your brand as a good choice for the asker's need, not just listing it. This is where scorers disagree most, so be explicit.

Our rule of thumb: would a busy buyer reading this answer come away thinking "I should look at Ledgerline"? If yes, it's a recommendation. If they'd come away with "Ledgerline exists", it's a mention.

Citations: yours, about you, and everyone else's

A citation is a source the answer links to. Engines show these differently. ChatGPT search shows inline citations and a Sources list (OpenAI Help). Google's AI features show supporting links. In the Gemini API, citations are attached to specific spans of the answer text (Grounding with Google Search).

Record every cited domain, not just yours, and classify each:

Our view: the "everyone else's" citations are the most underused data in GEO. Across a prompt set, the domains engines cite again and again are a ready-made outreach and PR list. Lesson 7 shows how to read them.

Ten edge cases

1. Cited, not mentioned. Your guide is a source, but only competitors are named. You're supplying the evidence for their recommendation. Score: cited Y, mentioned N. Flag it; it's one of the most fixable patterns.
2. Mentioned, wrongly. An outdated price, a feature you dropped, or the wrong target market. Score the mention, and mark accuracy as N with the wrong claim quoted.
3. Mentioned as an alternative. "If Tool B is too expensive, try Ledgerline." Usually a conditional recommendation.
4. Brand name is a common word. If your brand is "Ledger", make sure the answer means you and not accounting ledgers.
5. Named in the sources list only. The source title includes your name, but the answer text doesn't. Treat this as a citation, not a mention.
6. Recommended for the wrong use case. "Ledgerline is great for freelancers" when you sell to mid-market. Score it as recommended, inaccurate.
7. Parent company vs product. Decide once whether a mention of the parent counts.
8. The answer refuses or hedges. "I can't recommend specific vendors." Score all three as N, and note it. These answers count in your denominator.
9. A follow-up question instead of an answer. "What ERP do you use?" Score the first response as it is. Don't answer the follow-up during measurement; that changes the prompt.
10. Same brand, two spellings. "LedgerLine" vs "Ledgerline". Match case-insensitively and list known variants.

The scoring sheet

Use one row per answer. You'll reuse this sheet in Lessons 4 to 7.

From this you calculate three rates, each as a share of all answers scored (including refusals):

• Mention rate = answers with mentioned = Y ÷ all answers
• Recommendation rate = answers with recommended = Y or Y-conditional ÷ all answers
• Citation rate = answers with own_cited = Y ÷ all answers

See the glossary for mention rate and recommendation rate.

From MentionX data: how the outcomes come apart

Method. We ran read-only aggregate queries over completed MentionX engine runs from 22 August to 29 October 2026 (UTC): 2,154 runs across 18 projects, six engines via API with web search or grounding on. "Mentioned" means our extraction found the tracked brand in the answer text. "Cited about the brand" means at least one returned citation was linked by our extraction to the tracked brand. That covers its own site and third-party pages about it. Brands are anonymised.

Position, when it appeared. We looked at the 443 answers to unbranded prompts that named the tracked brand. In 341 of them (77%), the brand wasn't part of a ranked or numbered list our extraction detected; it appeared in prose. In the 102 where it was in a list:

How to read it. Mentions far outnumber citations about the brand, so a citation-only tracker would have missed most of these brands' visibility. When the brand made a list, it was usually low on it.

Caveats. The data covers 18 projects, mostly B2B software, and the engine mix is uneven. Lists can repeat a brand in sub-sections, and we used the first position. Our extraction is automated and makes mistakes. API answers may differ from consumer apps.

Worked example: scoring three Ledgerline answers

Illustrative example. Ledgerline and the answers below are invented. Tool A, Tool B and Tool C stand in for competitors.

Answer 1 (prompt: "best AP automation for NetSuite"): "For NetSuite users, Tool A and Ledgerline are strong options. Tool A is the best fit for larger enterprises [1]. Ledgerline is often chosen by mid-sized companies for its native NetSuite sync [2]." [1] toola.example/enterprise [2] reviewsite.example/ledgerline

Answer 2 (prompt: "how to stop duplicate supplier payments"): "Use a tool with duplicate detection, such as Tool B or Tool C [1]. Best practice is to match vendor ID, amount and invoice date [2]." [1] comparisonblog.example/ap-tools [2] ledgerline.example/blog/duplicate-payments

Answer 3 (prompt: "alternatives to Tool B"): "Alternatives include Tool A, Tool C and Ledgerline. Ledgerline starts at $99 per month." (no sources; Ledgerline's real entry price is $249)

Three answers, three different problems. Answer 1 is a win to protect: the review site carries it, so keep that profile current. Answer 2 is the "cited for someone else" pattern. Ledgerline's blog explains the method, but the answer recommends Tool B and Tool C. The fix is to name Ledgerline's own duplicate detection in that post. Answer 3 is an accuracy problem with no citation to correct. That usually means the wrong price lives in the model's training or in pages it didn't cite. Publishing clear, current pricing and updating third-party listings is the slow fix.

A single "visibility score" would have told Ledgerline it was visible in 2 of 3 answers and left it at that.

Which outcome to prioritise

Track all three, but choose a headline metric per prompt intent:

Our view: if you can only report one number to leadership, report the recommendation rate on unbranded compare and best-of prompts. It's the closest to pipeline, and the hardest to inflate.

Exercise: score ten answers

1. Take five prompts from your Lesson 2 exercise. Run each in two engines, logged out, to get ten answers.
2. Score them with the sheet above.
3. Find one answer where the outcomes disagree, for example cited but not recommended. Write one sentence on why you think it happened.
4. If you can, have a colleague score the same ten without seeing your scores, and compare the recommendation column.

Quiz

1. "Options include Ledgerline, Tool B and Tool C." How do you score it? Mention only.

2. Your guide is a cited source, but only competitors are named. What is this? Cited, not mentioned. Your content is supplying evidence for someone else.

3. Should refusals ("I can't recommend vendors") be left out of your rates? No. They count in the denominator.

4. In MentionX data, which was more common: mentioned without a citation about the brand, or cited without being mentioned? Mentioned without a citation (26% vs 1%).

5. What's the fastest way to find vague scoring rules? Have two people score the same answers independently and compare.

Next: Lesson 4: Your first manual visibility check.

Sources

• OpenAI Help Center, ChatGPT search
• Google AI for Developers, Grounding with Google Search
• Google Search Central, AI features and your website
• MentionX aggregate scan data, 22 August to 29 October 2026 (UTC), 2,154 engine runs, 18 projects. Anonymised and read-only.