You searched your own category in ChatGPT, and there it was: a competitor, cited as the source, while your page sat somewhere no one will ever see. That stings. If you have been asking why competitor gets cited by AI while you are left out, here is the good news: it is rarely random. Citation follows patterns you can read, and this guide shows you how to reverse engineer AI citations one page at a time. Treat it as a competitor AEO analysis you run on real pages, not a mystery you stew over.
Here is the reframe that makes this doable. Rank is not the answer. As of early 2026, only about 38% of Google AI Overview citations came from the top 10 organic results, down from 76% a year earlier, and roughly 88% of Google AI Mode citations come from outside the organic top 10. A page ranking 28th can beat a page ranking 4th. So this is not about outranking anyone. It is about running a structured teardown on the exact pages an engine cites, finding what they do that yours does not, and turning those gaps into a short fix list.
By the end, you will have a repeatable competitor AEO analysis you can run for any topic, on any engine, in an afternoon. Let's take it one step at a time.
Step 1: Map the buyer prompts your competitor is winning
Start with the questions, not the pages. Pull 10 to 15 buyer prompts where you already know the competitor shows up and you do not.
You have two sources for this. First, engine-side observation: run each prompt in ChatGPT, Perplexity, Gemini, Claude, and Google AI Mode in a private or incognito session, so your history does not skew the answer. Second, your own tracking, if you already watch prompts at the page level.
Stage each prompt by funnel: awareness, consideration, decision, or support. Then add a prompt-type column: informational, commercial, comparison, local, problem-solving, or transactional.
Done when: you have a one-page table with prompt text, funnel stage, prompt type, and a column for which engine currently surfaces the competitor.
Where do people go wrong here? They pick prompts that are too generic. "What is a CRM" tells you almost nothing. AI citations concentrate on commercial and comparison-shaped questions, like "best CRM for SaaS startups under 50 seats." Those are the prompts where the pattern is easiest to read, so weight your list toward them.
Do not skip the incognito part either. If you run these prompts in a logged-in session, your own history nudges the answer and you end up studying a distorted picture. A clean session gives you the answer a first-time buyer would actually see, which is the one that matters.
Step 2: Capture every URL the engine cites
Now go page by page. For each prompt, on each engine, record four things: the cited URL, the source domain, the source type (blog post, listicle, comparison, research report, product page, FAQ, glossary, forum, or video), and the exact sentence the URL was cited for. Screenshot every answer so you have a record.
The single biggest mistake in this step is treating "the competitor" as one URL. AI citations are page-level, not domain-level. The same competitor might win one prompt with a listicle, another with a research report, and a third with a product page. You are not studying a brand. You are studying specific pages.
One more distinction worth tracking: citation rate (when the engine links your page as a source) and mention rate (when it names you without a link) are different signals. A page can be mentioned and never cited. Note both.
This is the step where starting from real data saves you hours. DeepSmith's competitor citations view shows who wins citations for your prompts, on which exact pages, and how each competitor performs by engine. Instead of guessing which URLs to tear down, you begin the teardown with the real list already in front of you.
Done when: every row in your table has at least one cited URL per engine, each with the passage it was cited for.
Step 3: Score each cited page against a five-layer rubric
Here is the centerpiece. For every competitor URL you captured, score five layers from 0 to 2, for a maximum of 10. The total is that page's citation fitness. Do not overthink each score. A quick, honest read is enough.
| Layer | What to check | Score 0 | Score 1 | Score 2 |
|---|---|---|---|---|
| Prompt | Does the page match the prompt's intent and stage? | Mismatch or wrong stage | Partial match | Direct match on intent and funnel stage |
| Source | Is the format one engines tend to cite for this prompt? | Format mismatch | Partial fit | Clean fit (a true listicle for a "best X" prompt) |
| Content | Does the page carry extractability markers (descriptive headings, a 40 to 60 word opening answer, a table or definition list, an FAQ block, original data, a primary-source citation)? | 0 to 1 markers | 2 to 3 markers | 4 or more markers |
| Entity | Is the entity legible (named author with bio, Organization or Product schema, consistent naming, off-site presence)? | Anonymous and unstructured | Some schema or bio, inconsistent | Full schema, named author, consistent off-site graph |
| Authority | Is there external validation (tier-one mentions, peer backlinks, directory inclusion, third-party coverage)? | Self-referential only | Some external mentions, not tier-one | Tier-one placements plus steady third-party coverage |
Read the total like this. A page scoring 8 to 10 is a strong cited candidate, and your job is to catch up. A 5 to 7 means the fix list will be mostly structure and entity work. A 0 to 4 means the page is structurally invisible, so structure comes first.
Common mistake: scoring on word count. Length is not the lever here. AI Overviews cite pages under 1,000 words at majority rates. A tight, well-structured 700-word page can beat a sprawling 3,000-word one. Structure beats length, every time.
The trap underneath all of this is confusing what looks good to a human with what an engine can actually extract. A page can be beautifully designed and still be invisible to a model: bad heading nesting, missing schema, primary text hidden behind tabs, or content that only renders with JavaScript. Score for the machine, not the eye.
If format is where you feel weakest, that is the fastest layer to act on. Listicles alone account for roughly 45.8% of all classifiable AI citations, about 2.4 times the rate of reviews and several times the rate of a typical informational blog post. FAQs, how-tos, comparison tables, and dated data-rich pages all earn citations because each one hands the engine a clean, liftable unit.
Here is the quick version of which formats earn citations and why:
- Listicles: numbered, named items are trivial for an engine to extract.
- FAQs: each question and answer is a discrete unit a model can lift whole.
- How-tos: numbered steps read cleanly as a sequence, not buried in prose.
- Comparisons: a side-by-side table answers an "X vs Y" prompt in one glance.
- Data-rich research: engines quote specific statistics, especially dated ones.
- Glossaries: one tight definition per term is easy to pull.
Match the format to the prompt shape, and you remove the most common reason a good page never gets cited.
Pro tip: score each layer independently, then sanity-check the total. If a page scores high but is not cited, the prompt is probably answered from the model's training data rather than live retrieval. If it scores low but is cited, the engine is leaning on a source it already trusts. Both readings tell you where to spend your effort.
Step 4: Run the same teardown on your own page
Now turn the rubric on yourself. Pick the page you would most want cited for that prompt, and score it exactly the same way. Then compute the delta, layer by layer. Those deltas are your fix list.
Be honest here, and be specific. The right comparison is not your domain against their domain. It is their winning page for this prompt against your page for this prompt. Compare like for like, or the numbers will mislead you.
If you have never scored your own content this way, it can feel exposing. That is normal. Almost everyone finds a page they were proud of scoring a 4. That gap is not failure. It is the clearest to-do list you will get all quarter. This side-by-side is where the vague worry, why AI cites competitor not me, finally turns into a concrete list of gaps you can close. Running a full pass first, an audit before you rewrite anything, keeps you from editing the wrong pages.
Done when: for each prompt you have a competitor score, your score, and a per-layer delta.
Step 5: Cross-check the three authority signals
Scoring alone misses one thing: the authority layer is the hardest to read and the hardest to game. When two pages look identical on structure, this is often the real reason why AI cites competitor not me. So validate the authority layer directly with three checks.
Earned authority. Does the cited page have placements the engines already trust, like Forbes, Harvard Business Review, TechCrunch, or the standard publication in your vertical? Coverage the brand did not write and did not pay for counts most. Self-asserted authority does not.
Entity clarity. Is the identity machine-readable? Look for Person schema on the author, Organization schema for the company, Product schema where it fits, and consistent naming across the page, the About page, LinkedIn, Crunchbase, and Wikipedia if present.
Citation architecture. Are the extraction surfaces built in? FAQ blocks, comparison tables, answer-first structure, published and modified dates in the metadata, and primary-source citations inside the body.
Most winning pages win on at least two of these three. And here is why this step matters so much: a page that scores a perfect 10 on prompt and format fit can still lose to a page that scores a 6, if that lower page has stronger authority signals. Do not treat Google rank as a proxy either. Remember, most AI citations come from outside the top 10, so a page ranking 4th can lose to a page ranking 28th with better citation architecture.
Done when: each cited URL has a one-line note on whether it wins on earned authority, entity clarity, or citation architecture.
Step 6: Turn the gaps into a prioritized fix list
You now have deltas. Time to make them a plan you can ship. Sort every gap into one of four buckets, and give each item an owner, a deadline, and a test.
- Structure fixes. Rewrite headings as questions or named concepts, add a 40 to 60 word opener, add tables and definition lists, add an FAQ block, and mark it up with HowTo or FAQPage schema.
- Content fixes. Add original data, cite primary sources, add an expert quote with a real name attached, and refresh statistics with dates.
- Entity fixes. Add a named author with a bio and credentials, add Organization and Person schema, clean up your LinkedIn and Crunchbase presence, and standardize how you name your brand and products.
- Authority fixes. Pitch guest posts and features to tier-one publications, earn inclusion in trusted roundups, and build founder presence on LinkedIn and YouTube.
Done when: you have 8 to 20 items, each tied to a specific page's gap, each with an owner, a deadline, and a verifiable test.
The mistake to avoid is treating this as a wishlist. Every line should answer one question: "if we ship this, the page moves from a score of X to a score of Y." A fix with no test is a hope, not a task. Without that number, your priorities collapse into whatever feels urgent.
This is also the point where the teardown turns into published work, and that is usually where lean teams stall. DeepSmith's writing pipeline produces on-brand articles with the answer-first structure, headings, internal links, schema, and metadata built in during creation, so the structure fixes you just scoped ship as finished pages instead of sitting on a to-do list. You keep the editorial judgment. The manual formatting work goes away.
Step 7: Ship, wait one cycle, and re-test
You shipped. Now resist the urge to refresh the answer every hour. AI retrieval runs on cycles, and they take time.
Wait one full cycle before you judge anything: roughly 1 to 2 weeks for Google AI Mode and AI Overviews, and 1 to 4 weeks for ChatGPT and Perplexity, where the timing is less predictable. Then re-run your original prompt set on each engine, screenshot the responses, and record the change in answer share. Any prompt where you still lose goes back to Step 4 for a second pass.
Here is the part that trips people up. Some prompts will not move, no matter how good your fixes are, because the model answered from its training data instead of fetching live sources. That is not your fix failing. It is the engine not retrieving. This is exactly why you track the outcome per prompt, not an average visibility score. Learning to diagnose lost AI citation results prompt by prompt is what keeps you from chasing a number that hides the real story.
Tracking prompts on a schedule turns this from a manual chore into a system. DeepSmith checks your tracked prompts automatically and reports mention rate, citation rate, and share of voice over time, so the re-test in Step 7 happens without you rerunning every prompt by hand.
Done when: the same 10 to 15 prompts have been re-tested, the answer-share delta is recorded, and any unresolved prompts are back in the loop.
Your teardown, on repeat
That is the whole method. Map the prompts, capture the cited pages, score them on five layers, score yourself, cross-check authority, ship a tested fix list, and re-test after one cycle. It is a loop, not a one-time project. Each pass is a fresh competitor AEO analysis, and each one gets faster because your prompt table and rubric are already built. Run it on your highest-value prompts first, then widen out as you build momentum.
Once you can reverse engineer AI citations on demand, the answer to why competitor gets cited by AI stops being a source of dread and becomes a checklist. You do not need to fix everything this week. Pick the one prompt that matters most, tear down the page that beats you, and ship the top three fixes. That is a real start, and it is more than most of your competitors are doing.
Want to skip the manual URL hunting and start your teardown from the real list of pages an engine cites for your prompts? Start a free DeepSmith trial and see which competitor pages win your citations, then produce the content to close the gap.
