Fixing Low AI Citation Frequency
Low citation frequency is the in-between problem. Assistants can reach your site and do cite it, but only for a few questions, or only on some runs, while competitors turn up far more often. Access is not the issue here; choice is. When an assistant has several good pages to pick from, yours is not the one it picks.
This guide shows how to measure frequency in a way that is not fooled by random variation, how to work out why you lose particular prompts, and what kinds of change give an assistant a genuine reason to cite you more.
Why citation frequency moves around
Ask an assistant the same question five times and you may see five slightly different lists of sources. That is not a fault in your measurement; it is how these systems work. Each answer is generated fresh. The search step can return different pages, the model can choose different passages from them, and the products are updated without notice. Location, language, the wording of the question and whether the user is signed in can all change the result too.
So a single screenshot tells you almost nothing about frequency. If you were cited once, you might be cited half the time or one time in twenty. If you were missing once, you might still be a regular source. The only useful view is a rate: across a fixed set of questions, asked repeatedly, how often does each assistant cite you?
That also means small changes in the rate from one week to the next are usually noise. Look for movement that holds across several measurements, and across many prompts, before crediting or blaming any change you made.
Measuring it properly
Keep your measurement simple enough to repeat every few weeks without effort. A spreadsheet with one row per prompt and assistant is enough. These are the figures worth working out from your results:
| Figure | How to work it out | What it tells you |
|---|---|---|
| Citation rate | Runs where one of your pages is linked, divided by all runs of that prompt. | How reliably you are chosen as a source for that question. |
| Mention rate | Runs where your business is named, linked or not, divided by all runs. | Whether the assistant knows you are relevant, even when it links elsewhere. |
| Share against competitors | Your citations divided by all citations to you and your named competitors on the same prompts. | Whether you are gaining or losing ground, independent of how many sources each answer shows. |
| Cited page | The URL on your site that was linked. | Which pages are doing the work, and whether the right page is cited. |
A high mention rate with a low citation rate is a specific pattern: the assistant knows your business but prefers other pages as evidence. That usually means your pages do not answer the question as directly as the pages it links, which is fixable on your own site.
Reading the sources you lose to
For every prompt you lose, open the pages that were cited instead. The kind of page that wins tells you what the assistant was looking for.
| What gets cited instead | What it suggests | Your response |
|---|---|---|
| A competitor's page on exactly that question | They have a dedicated answer; you have a general page. | Write or restructure a page around that question. |
| A comparison or "best of" article on another site | The assistant wants a spread of options, not one seller's view. | Make sure you are included in those comparisons; see the brand mentions guide. |
| A government, regulator or trade body page | The question needs an authoritative fact such as a rule, limit or standard. | Cite that source on your page and add what it does not cover: how it applies in practice. |
| A forum or community thread | The assistant wants first-hand experience. | Add real examples, case notes and specifics from your own work. |
| A page with a table of prices, specs or figures | The question needs data that can be quoted. | Publish your own figures clearly, with dates and conditions. |
| Your own page, but a weaker one | The assistant found you, but the wrong URL. | Strengthen the right page and link to it clearly from the weaker one. |
The fix, step by step
Step 1
Build a fixed prompt set and baseline
Choose 30 to 60 questions a potential customer would ask before they know your name, grouped by intent: definitions, how-to, costs, comparisons and local questions. Write them down exactly and do not change the wording between rounds, or the results stop being comparable. Run the set through the AI Visibility Tracker and record, for each question and assistant, whether you were cited, mentioned or missing, which of your pages was linked and which competitors appeared.
Step 2
Sort prompts into won, shared and lost
Mark a prompt as won when you are cited in most runs, shared when you are cited in some, and lost when you are rarely or never cited. Shared prompts are the quickest gains, because the assistant already considers your page and only needs a stronger reason to choose it. Lost prompts need more work. Sort each group by how close the question is to a purchase, and start with the commercial questions rather than the easiest ones.
Step 3
Study the sources cited on lost prompts
For each lost or shared prompt, open every page the assistant cited instead of you. Note what each gives the reader that yours does not: a direct answer in the first lines, a specific figure, a table, a named expert, a recent date, or first-hand experience. Also note the type of site, because a question answered by comparison articles needs a different response from one answered by a regulator. Keep these notes next to the prompt.
Step 4
Give each lost prompt a page that answers it
Check that each lost question has one page on your site whose main purpose is to answer it. If the answer is buried inside a long general page, give it its own clearly headed section with a direct answer at the top, or a dedicated page when the topic is big enough. Avoid creating many thin pages for small wording variations of one question; one thorough page that covers the variations is usually a stronger candidate.
Step 5
Add something only you can publish
Pages that repeat what every competitor says give an assistant no reason to prefer them. Add information you hold and others do not: your own prices and how they are worked out, typical timescales from real jobs, results of tests you ran, anonymised examples, photos with captions, or figures from your own records with the date range stated. Explain how you collected any data. Never publish invented or estimated numbers as if they were measured.
Step 6
Cover the follow-up questions around each topic
People and assistants rarely stop at one question. Google has described AI Mode breaking a question into several related searches, and other assistants also search more than once for complex questions. List the natural follow-ups for each target topic, such as costs, risks, alternatives, how long it takes and who it suits, and answer them on the same page or on closely linked pages, so your site can supply more than one part of the answer.
Step 7
Keep key pages current and dated honestly
For questions where facts change, such as prices, rules, product ranges or opening times,
stale pages lose to current ones. Review your target pages on a schedule, update the facts that have
changed and show a visible last updated date only when the content really changed. Put the same date
in dateModified in your Article markup. Changing the date without changing the page misleads
readers and does not make the content any more useful.
Step 8
Re-measure on the same prompts
Wait until the changed pages have been recrawled, which you can check in Search Console and the access log, then run the same prompt set again. Compare citation rate, mention rate and share for each group of prompts, not individual answers. Record which pages you changed and when, so you can link movement to specific work. Expect gains to be uneven between assistants, and repeat the cycle every one to three months.
Being the source other sources use
A useful way to think about frequency is this: when several pages say the same thing, an assistant can cite any of them. When one page is where a fact comes from, it has a stronger claim. You do not need a research department to be that page. Small businesses hold plenty of first-hand information that nobody else publishes.
- Your own pricing, explained. A page that says what a typical job costs, what makes it cost more, and what is included is quoted far more readily than "prices vary, contact us".
- Numbers from your own records. "Of the 140 boiler services we carried out in 2025, the most common fault was..." is original, as long as it is true and the period is stated.
- Tests and comparisons you ran. If you sell or fit products, a fair comparison of the ones you have used, with the criteria stated, is information an assistant cannot get elsewhere.
- Clear definitions of trade terms. A plain definition, written by someone who uses the term daily, is often the passage an assistant needs for a "what is" question.
When frequency is low for other reasons
The assistant is not searching
For some questions, especially simple or general ones, an assistant may answer from its trained knowledge without running a search. No sources are shown, so nobody is cited. These prompts will look like losses in your data but are not really winnable through your pages. Mark them and leave them out of the rate.
Your brand name is ambiguous
If another company shares or nearly shares your name, the assistant may mention "you" while linking
to them, or the reverse. Consistent naming, a clear about page and Organization markup with sameAs
links help separate the two. The wrong
attribution guide covers this in detail.
Local questions
For "near me" and town-name questions, assistants often lean on map listings and review sites. A complete, accurate Google Business Profile and Bing Places listing, plus consistent details on local directories, can matter as much as your own pages here.
One assistant is far lower than the rest
If you are cited often in one assistant and hardly at all in another, check access and indexing for the weak one before changing content. A firewall rule that challenges one crawler, or a site that is missing from Bing, produces exactly this pattern.
Holding on to the gains
- Keep the prompt set stable. Add new questions as a separate group rather than changing the old ones, so trends stay comparable.
- Assign an owner to each key page. Someone should check its facts on a schedule, not only when a customer complains.
- Watch competitors on won prompts too. A new, better page from a competitor can take a won prompt back to shared within weeks.
- Record every change. A simple log of what changed on which page, and when, is what lets you learn which kind of work moves the rate for your market.
If you are not being cited at all, start with the guide to getting cited by AI search, which covers access, snippet controls and indexing.
1. What is a good AI citation frequency?
There is no published benchmark, and any single figure would depend heavily on your market and the questions you track. A better yardstick is your own trend and your share compared with the competitors who appear on the same prompts. If your share of citations on commercial questions is rising over several measurements, the work is paying off. Be cautious of anyone quoting a universal target rate, because the assistants do not publish such figures.
2. Why was I cited yesterday but not today?
Because AI assistants generate each answer fresh. The search they run can return a slightly different set of pages, the model can pick different passages, and the products themselves change over time. Being cited on one run and not the next is normal, particularly on prompts where several pages answer the question about equally well. That is exactly why frequency should be measured as a rate across repeated runs rather than judged from one answer.
3. How many times should I run each prompt?
Enough to see a pattern rather than a single outcome. Running each prompt several times per measurement round, and repeating the round every few weeks, gives a far more reliable picture than one run. What matters most is consistency: use the same wording, the same assistants and a similar schedule each time. If you change how you measure, start a new baseline rather than comparing the new figures directly with the old ones.
4. Why does ChatGPT mention my business but link to other sites?
A mention without a citation usually means the assistant recognises your business as relevant but finds a better page elsewhere to support what it is saying. That page might answer the exact question more directly, contain a specific figure or table, or be a comparison article that lists several options. Open the linked pages, see what they provide, and add a clear, direct answer with the same kind of detail to the matching page on your own site.
5. Does publishing more content increase citation frequency?
Not by itself. Many near-identical pages, one for each variation of a question, tend to compete with each other and can look like doorway pages to search engines. What helps is having one strong page for each real question, covering the natural follow-up questions, with information that is specific and ideally original. A smaller number of thorough, well-maintained pages is usually a better candidate for citation than a large number of thin ones.
6. Do updated dates help pages get cited more often?
Freshness matters for questions where the facts change, such as prices, rules or product ranges, and assistants often favour current information for those. But the benefit comes from the content being current, not from the date itself. Update the date only when you have actually changed the page, and keep the visible date and the dateModified value in your markup the same. Changing dates without changing content misleads readers and does not make the page more useful.
7. How long before changes show up in citation frequency?
First the changed page has to be recrawled by the search systems the assistants use, which can take from days to several weeks. Then the assistant has to start preferring it, which may take several measurement rounds to show clearly. Plan to measure the same prompt set at least two or three times after a change before drawing conclusions, and compare groups of prompts rather than individual answers, because single results vary too much.
The AI Visibility Tracker asks AI assistants the questions your customers ask and records whether your brand is cited, mentioned, recommended or missing, so you can see which prompts you are losing and track progress over time.
Open AEO Fixes →