GEO competitor analysis: how often rivals show up in AI answers and where the citations come from
1. Pick competitors: 3 to 5, no more than 5
A longer competitor list isn’t better. Past 5, the table gets too wide to read once you run the question set, and nobody records it carefully. 3 to 5 direct competitors in the same category, same market and same price range is enough for a benchmark table you can actually draw conclusions from.
| Competitor type | Include? | Reason |
|---|---|---|
| Direct competitor, same category and price range | Required | AI will compare you head-to-head on similar questions |
| Same category, different price range | Pick 1 | See whether AI places you in the wrong price tier |
| Different category, but customers compare you | Pick 1 | See whether AI writes alternatives into the answer |
| Industry leader | Exclude | Too different in scale; the benchmark gives you nothing actionable |
| Small player that just entered the market | Exclude | AI hasn’t indexed them yet, so you get zero either way |
There are two easy mistakes when picking competitors. First, benchmarking against the industry leader, only to find they get cited every time and you can never catch up, which tanks team morale. Second, picking only the competitors you already know and missing the small players that show up often in AI answers but never crossed your radar. The second is more dangerous: the competitors your customers see in AI may not be the same ones your sales team talks about.
A real-world example: say you make industrial valves, and your sales team sees A, B and C as the competitors. But after running the question set, you find that on the long-tail question “how do I troubleshoot pressure loss in a butterfly valve,” a company called D you’ve never heard of gets cited every time, because an engineer has been writing technical posts about it on Reddit for 5 years. That’s the value of competitor analysis: the competitor list in AI’s eyes may not match the list in your sales CRM.
How do you find these “AI rivals”? Put 5 open-ended questions in your question set, like “Which butterfly valve brands are worth recommending?” or “How do I choose a PN16 butterfly valve manufacturer?” The 5 to 8 brands that show up in the AI answers form your full candidate pool. Pick the 3 to 5 that appear most often and add them to your formal benchmark table.
2. Run the question set: same questions, same engines, same time window
Competitor analysis and self-monitoring use the same question set, but how you run it matters. You have to run it in the same week, in the same time window, across all three engines. Otherwise engine version updates and the randomness of AI answers will contaminate the comparison.
| Field | Description | Example |
|---|---|---|
| Original question | The exact wording from your fixed question set | “Maximum operating temperature of a PN16 butterfly valve” |
| Engine | ChatGPT / Gemini / Perplexity | Perplexity |
| Brand mentions | Who appears in the answer text | You, A, C |
| Mention position | Which sentence it appears in | Sentence 2 |
| Cited links | Whether the official site is in the source list | You: yes, A: yes, C: no |
| Co-cited domains | Who are you cited alongside? | Wikipedia, industry associations, A’s official site |
Don’t just ask everything once and leave. Run the same question 3 times in the same engine and see whether the answer holds steady. AI answers are random, so a single run can be noise. If you run the same question 3 times and the order in which you and company A are mentioned keeps changing, that question isn’t a stable gap right now, so don’t draw conclusions from it.
This shares the same difficulty as Monthly monitoring of a fixed question set It uses the same infrastructure. The difference is: self-monitoring only fills in your own column, while competitor analysis fills in the columns for 3 to 5 rivals as well. For choosing a monitoring tool, see AI engine citation monitoring dashboard.
3. Break down source domains: the gap is off-site, not on-site
This is the most counterintuitive point in the whole article. Most people finish a competitor analysis and go back to rewrite their own website copy, titles, and product pages. But you’ll find that a rival gets cited more by AI not because their website is better written than yours, but because they’re mentioned over and over on third-party sites.
| Source domain type | AI citation weight | Rival A’s share | Your share | Gap assessment |
|---|---|---|---|---|
| Industry associations and standards bodies | Extremely high | 35% | 5% | Major gap |
| Industry media and top blogs | High | 25% | 15% | Moderate gap |
| Q&A platforms (Reddit/Zhihu) | Medium | 15% | 10% | Close |
| Mutual citations between competitor sites | Medium-low | 10% | 30% | You’re higher, but low quality |
| Directory sites and spam sites | Low | 15% | 40% | You need to clean these up |
One look at this table tells you: 35% of rival A’s citations come from industry associations and standards bodies, domains AI inherently trusts. 40% of your citations are directory sites, which AI gives very little weight. You could rewrite your website copy 10 times and still not close that 30-percentage-point gap in authoritative sources.
So the next step in GEO competitor analysis isn’t to go back to your website and rewrite copy. It’s to look at industry association sites, trade media, and the top Q&A platforms to see how Competitor A gets mentioned. Is there a signed expert article? A product comparison review? Engineers discussing it organically on a forum? List all of these — that’s the gap you need to fill.Off-site factual sources That article went deep on the role of industry directories, encyclopedias, and B2B platforms in AI answers.
A scenario: after benchmarking, you find Competitor A is listed as a recommended supplier in an industry association’s technical white paper. AI reads that source every time, so A gets cited in every question like “recommended suppliers under XX standard.” What you need to do isn’t write “we’re also a recommended supplier” on your own site — it’s contact that association and see if you can submit your product data to be included in the next white paper.
4. Output a benchmark table: one table that makes the gaps clear
Competitor analysis needs to produce a one-pager, or the 2 person-days of work are wasted. The benchmark table structure is simple: you and 3 competitors, five metric layers across, question buckets down.
| Metrics | You | Competitor A | Competitor B | Competitor C |
|---|---|---|---|---|
| Category term mention rate | 20% | 65% | 40% | 15% |
| Comparison term mention rate | 10% | 50% | 30% | 20% |
| Citation link hit rate | 15% | 55% | 25% | 10% |
| Authoritative source share (Tier A domains) | 20% | 60% | 35% | 15% |
| Number of inaccurate descriptions | 1 | 0 | 2 | 0 |
| Number of co-cited top domains | 3 | 12 | 6 | 2 |
This table tells you three things directly. First, you’re 45 percentage points behind Competitor A on category terms — that’s the biggest gap and the top priority. Second, you have 1 inaccurate description that must go through the AI got your brand wrong correction process. Third, you only have 3 co-cited top domains versus Competitor A’s 12, meaning A’s presence density on third-party authoritative sites is 4 times yours. That’s where your next backlink and PR efforts should go.
Run a full benchmark once a quarter; monthly, just watch the trend. Don’t re-benchmark every week — AI answers don’t fluctuate much month to month, and weekly runs will trap you in meaningless number swings. A full benchmark takes 2 person-days, monthly tracking takes 0.5 person-days. That’s a sustainable rhythm.
5. Pitfalls: the three most common misjudgments
Competitor analysis is most prone to three types of errors, and each one will lead your team astray.
| Misjudgment | Performance | Consequence | Fix |
|---|---|---|---|
| Drawing conclusions from a single run | One run and you declare the competitor the winner | Mistaking randomness for a trend | Run the same question 3 times, take the median |
| Only comparing mention counts | Whoever gets mentioned more is stronger | Ignoring source quality | Break down co-cited domains by tier |
| Only looking at your own website content | Go back and rewrite your website copy | Can’t close off-site gaps | Break down source domains first, then decide what to do |
The fourth mistake is treating “being mentioned” and “being recommended” as the same thing. When AI mentions you, it might be holding you up as a cautionary example, or it might just be listing you alongside others. So your benchmark table must have a column for “type of mention”: positive recommendation, neutral listing, negative description, inaccurate description. That column is often more instructive than the mention count itself.
One last point: competitor analysis isn’t about proving you’re better than your rivals. It’s about finding the gap to close next quarter. If you finish reviewing the benchmark table and can’t name a concrete action like “next quarter we need to show up in an industry association white paper,” the analysis was wasted.Competitor content gaps That article covered gap analysis from a traditional SEO angle. The difference with a GEO lens is this: most of the gap lies in off-site sources of truth, not on-site content.
FAQ
How many competitors should you include in a GEO competitor analysis?
Why do competitors get cited by AI more often, if it’s not because their website is well written?
How often should you run competitor analysis?
What if I run the same question 3 times and get different results each time?
What fields should the benchmark table output?
My competitor analysis found AI grouping me with the wrong competitors. What should I do?
Final thoughts
The biggest difference between GEO competitor analysis and traditional SEO competitor analysis is that it doesn’t look at competitors’ rankings and traffic — it looks at who gets cited for them in AI answers. Shift that perspective and you’ll find that most competitors’ moat isn’t on their own website, it’s the density of third-party authoritative sources.
Three steps to put it to work: this week, list 5 open-ended questions and run them once to map out the full list of competitors in AI’s eyes; next week, run a full benchmark with 20 fixed questions across three engines and sort the cited domains into four tiers; in the third week, produce a one-page benchmark table and decide which authoritative source gap to close next quarter.
One line worth remembering: GEO competitor analysis isn’t about what competitors say, it’s about who speaks for them — the density of off-site authoritative sources is the real moat.