How to Track AI Citations With 11 Daily Checks
Google Search showed 37,700 generative AI citation impressions for our site over three months.
That number became useful when we logged the questions, engines, brands and cited URLs behind it each day. This guide gives you the 11 checks we use to do that.
Ahrefs ran 15,000 prompts through ChatGPT, Gemini and Copilot. Only 12% of the links cited by those engines appeared in Google's top 10 for the same prompt. For ChatGPT, the figure was 2.1%.
A normal rank tracker misses most of that activity. You need a separate record of which brands and pages appear inside AI answers.
If you want one place to research the gaps, publish the missing pages and manage distribution, try Distribb.
Set up the daily files
The source workflow uses 11 Markdown files inside .claude/agents/. Each file holds one job. A scheduled command runs the files and appends the results to logs.
For example:
0 7 * * * cd ~/geo && claude -p "$(cat loops/01.md)" >> logs/01.log
Run the first checks a few minutes apart so they do not all start at once. Keep every dated result. Overwriting yesterday's file removes the comparison that makes the system useful.
1. Track citations across several runs
Run every buyer prompt on each engine three times. Record the prompt, engine, run number, every brand named, every URL cited and the position of your first mention.
Merge the results into citations.csv under the current date. Then calculate your share of voice and compare it with the previous day.
Repeated runs matter because the answers move. A Washington State test ran identical prompts 10 times and received matching answers only 73% of the time. One monthly screenshot can catch an answer that disappears on the next run.
2. Measure prompt coverage
The citation file counts brands and links. The coverage file counts questions.
For every prompt in prompts.csv, store a simple yes or no for whether your brand appeared. Also record the brands named in your place.
Sort the missing questions by buying intent. A question about price, alternatives or a direct comparison usually deserves attention before a broad informational query.
3. Map the sources helping competitors
When an engine recommends a competitor, collect every URL cited in that answer.
Remove duplicate URLs. Rank each domain by the number of answers it influenced. For the top 20 pages, record whether your company could be added and who owns the page.
That turns a vague goal such as "get mentioned on Reddit" into a list of exact threads, roundups and review pages to work on.
4. Find weak answers you can improve
Write 30 buyer questions for each topic you cover. Run the questions through the engines and score every answer from 1 to 5.
Use four checks:
- Is the answer specific?
- Does it cite a source?
- Does it give a number?
- Does it hedge instead of answering?
Keep every answer scoring 2 or below. Each weak answer becomes a brief with the question, the answer the engine should have given and the evidence needed to prove it.
In the source run, 10 topic agents scored 300 questions and found 68 weak answers.
5. Check your llms.txt file
An llms.txt file is a Markdown list of the pages you want an AI system to understand.
Crawl your sitemap and split the site by section. Ask each agent to choose the pages a buyer or an AI engine would need, then write one description under 25 words for each URL.
Validate that every URL returns a 200 status. The same check can list pages missing Organization or Product schema. Run it weekly because the sitemap changes.
6. Audit the facts engines repeat about you
Put the verified facts about your company in facts.md.
Then ask five engines the same questions about your company: what it does, who it serves, what it costs, who competes with it and when it was founded.
Compare each answer with facts.md. Record the claim, the engine, whether it is true, false or outdated, and the page that could correct it.
A wrong price or competitor can send a buyer to the wrong product. Fix those claims before a minor date error.
7. Rank the outside sources that influence your category
Peec AI examined 30 million cited sources and found Reddit supplied roughly 40% of citations across the engines in its study. Reddit, Wikipedia, YouTube, LinkedIn, Forbes and 10 other domains covered about 68% of all cited sources.
Your own citation data should tell you which pages on those domains matter for your category.
Group the URLs in citations.csv by domain, then by subreddit or site section. Record how many tracked prompts each page influenced, whether your brand appears and what inclusion would require.
8. Alert on citation losses
Compare today's citation file with the previous seven days.
Create an alert when:
- A prompt cited you for at least three days and stops today.
- A new competitor enters the top three mentions.
- One of your URLs stops appearing as a cited source.
Include the last answer that cited you and the current answer. Send the alert to the place your team already checks each morning.
9. Repair pages after citation decay
For every lost URL, compare your page with the pages now cited in its place.
The rewrite brief should contain a 3 to 5 sentence summary that makes sense when quoted alone, the direct answer in the first paragraph, an FAQ built from real questions in coverage.csv and one statistic from your own data.
Use product analytics, billing history, support tickets or review data. Never invent a number to fill the space.
10. Write one daily visibility row
Read citations.csv, coverage.csv and the freshness alerts. Add one row to a sheet with:
- Share of voice for each engine
- Prompts covered out of the total tested
- New citation wins and losses
- The pages that need work today
Keep the report short and numeric. Its job is to decide today's work, so the action should be visible without opening the raw logs.
11. Publish into the gaps you found
The first 10 checks show which questions and outside pages influence the answers. Use the final check to publish the missing material.
Ahrefs tracked one million new URLs for a year and found that 1.74% reached Google's top 10. It also found that 72.9% of pages in the top 10 were more than three years old.
That makes established publishing accounts useful for some gaps. The source workflow includes Medium, Substack, YouTube and GitHub, depending on the format and the sites already cited for the topic.
For local or vertical searches, combine a parent keyword with the buyer type and location. Pull FAQ questions from Search Console with a question-word regex.
The nightly publishing log should reject a keyword already used on that account. Give each article a quotable summary, an FAQ based on real queries, one useful image with descriptive alt text, and internal links to related posts. Record failed publishes so the next run can resume them without creating a duplicate.
Which check would save you the most time this week: citation tracking, prompt coverage, freshness alerts or content decay?
Distribb can handle the research, publishing and distribution work behind this system.
