TL;DR: Find the pages the engine cites for the question you're losing, and check whether your competitor is in them. Usually they are, which turns a vague strategy problem into a short list of pages to get into.
Ask the engine the question. Write down every source it cites. Open each one and search for your competitor's name, then your own. Do this on each engine you care about. It takes about 10 minutes, and it sorts you into one of the six causes below.
You are absent from the pages the engine cites
When an engine searches before answering, it names the brands in the pages it retrieved. If those pages don't mention you, you're not in the answer, however good your own site is.
This is the one most people underestimate. In 10 of the 14 markets measured, 80% or more of the cited sources were somebody else's article and not any vendor's own site. In those markets, being named in other people's pages is what gets you into the answer. The research has the market-by-market numbers.
How to tell: look at what the engine actually cited. If the sources are roundups and guides, read them and look for your name.
Cost: weeks to months. It's outreach work more than publishing work.
Your competitor is in those pages and you are not
This is the same problem with a more specific fix. If the same three roundups keep getting cited and your competitor is in all three, the job is getting into those three pages. That's a finite, nameable task.
How to tell: list the cited sources across 10 of your questions and count how often each competitor appears.
Cost: weeks. The target list is short, so this is generally the highest-leverage work here.
The question is answered from memory and you weren't in the training data
Some questions never trigger a search. The model answers from what it learned, which closed months ago, and no page you publish now gets read.
How to tell: look for citations in the answer. No sources, a confident tone and a broad question usually means memory. Check a few times, since the same question varies.
Cost: months. The work is breadth of presence, like reviews, directories and consistent coverage.
You are described inconsistently
If your site, your G2 profile, your LinkedIn and three articles describe you four different ways, there's no stable fact for a model to learn or a retrieval step to match. Competitors with one boring, consistent sentence tend to win here.
How to tell: search your brand name and read the first 10 descriptions of what you do. If you can't sum them up in one sentence, a model can't either.
Cost: days of editing, then months for it to spread.
The engines can't read your site easily
If your robots.txt blocks the AI crawlers, your content renders entirely client-side, or key pages return errors to a non-browser request, you're invisible regardless of quality. This is the least common cause (2.9% of 818 cited domains block GPTBot, from the LLM SEO post), but when it's your problem, nothing else on this list works until it's fixed.
How to tell: fetch your own pages without JavaScript and see what comes back. Check robots.txt for GPTBot, OAI-SearchBot, ClaudeBot and PerplexityBot by name. A blanket allow for Googlebot doesn't cover them. The AI crawler checker does this in a minute.
Cost: days. It's the cheapest work available.
You genuinely are not the best answer
If the question is "best enterprise X" and you serve teams of 5, the engine naming an enterprise product is right.
How to tell: read the answer as a buyer would. If the recommendation is reasonable, the question is wrong for you.
Cost: none. Stop tracking that question and track the one your buyer actually asks.
Which one do you have
| Symptom | Likely cause |
|---|---|
| Engine cites sources, none mention you | Absent from cited pages |
| Engine cites sources, competitor in most | Competitor owns the cited set |
| Answers have no citations at all | Memory mode |
| You appear but are described wrongly | Inconsistent description |
| No AI engine cites you anywhere | Crawler access |
| The named competitor is a better fit | Wrong question |
If the engine cites your page but still doesn't name you, the page usually mentions you in passing while your competitor gets a paragraph, or describes you vaguely. Getting your entry on that page improved is an easier ask for an editor than getting added.
Check what kind of question it is
How specific the question is changes which pages get cited. In a Peak Answer study of 30 niches and 198 questions, ChatGPT cited the brand's own site for 35.6% of citations on general questions and 60.3% on specific ones. For Google organic it went from 30.5% to 49.9%. Publishers and directories nearly disappeared on specific questions. The full breakdown is in AEO vs SEO.
- General question: look at which lists and publishers are cited, and get into them.
- Specific question: most of the cited pages will be vendor sites, yours and your competitors'. The fix is usually a page on your own site that answers the question directly.
ChatGPT and Google also cite mostly different sites, so check each separately.
Run it across ten questions
One question tells you about one question. Ten gives you a plan, and it takes about an hour. For your 10 most important buying questions, on two engines, record every cited domain, then count. The output looks like this:
| Domain | Cited on | Competitor present? | You present? |
|---|---|---|---|
| g2.com | 7 of 10 | Yes | No |
| reddit.com | 6 of 10 | Yes | No |
| [industry publication] | 5 of 10 | Yes | No |
| [large publisher] | 4 of 10 | Yes | No |
Usually 3 to 5 domains account for most of the citations. That's your list. Sort it by how reachable each one is, best first:
- Review platforms: you claim a profile, no permission needed
- Forum threads: you answer a question, today
- Small newsletters and one-person publications: they need material
- Niche directories: a form
- Trade publications: a contributed article, if you have something to say
- Established category publishers: a relationship, measured in quarters
- Competitors' own sites: no
The most-cited domain is usually the least reachable one. Expect a low hit rate when you pitch. 2 out of 10 is a good result, and 2 pages in the cited set can change the answers.
Your first month
Week 1. Check robots.txt by crawler name. Fetch two pages without JavaScript. Run the ten-question tally and build the domain table. You'll have a target list, and you'll know which questions are in memory mode and which are broad or specific.
Week 2. Fix anything technical. Claim every review profile in the tally. Write to the three most reachable cited sources with a specific reason their page gets better, like a data point they don't have or a correction to something out of date. Here are a few pitch emails you can adapt.
Week 3. Write the one page you're missing: the question you lose where nobody has written a good answer. One page, not ten. For specific questions, this goes on your own site.
Week 4. Re-ask the 10 questions and record the result. Nothing will have moved yet, but you'll have a baseline to compare against in month 3. An AEO tracker can re-ask and record them for you.
Months 2 and 3. Repeat weeks 2 and 3.
FAQs
AI answers are generated fresh each time you ask, so some variation between asks is normal. Ask the same question several times and look at how often you appear.
A smaller competitor usually wins AI answers by being in the pages that get cited. Company size matters less than being named in those pages.
Being on page one of Google helps somewhat. Retrieval often pulls from search results, so ranking helps there. For answers that come from memory, it makes no difference.
Start by tracking 10 to 20 buying questions, measured weekly.