How to Run a Citation Gap Analysis

Sooner or later someone asks you for budget "to get us onto the sites ChatGPT cites." It is a fair request, and it usually arrives with a list.

Buyers do ask assistants. In Forrester's 2025 survey of business buyers, 94% used AI in their buying process, and buyers reported that AI tools help them "consider more vendors in less time."

The pages those answers read are now part of how your company is found, and most of them belong to someone else. The reports you already get cannot show them: in our research, 89.6% of the websites ChatGPT cited were not on Google's first page for the same buyer question.

A citation gap analysis finds those pages and sets them against where your company appears.

Finish the analysis as a work queue: each cited page gets a verdict, an owner, and a date. Rebuild it monthly because the answers move.

We ran the whole audit on email marketing software, a category crowded enough that every kind of website turns up. In our research, shortlist and evaluation questions went to ChatGPT and Google AI Mode in September 2026, and the cited pages of every site on the two resulting lists were opened on October 15.

You will build that queue with two prompts, then see how we handle the monthly work at Qvery.

One limit before the steps: it finds where your company is missing from the sources an answer cites. It does not tell you why an engine chose a source, and getting onto a page does not promise that the next answer names you.

What You'll Need

This builds the gap list, the queue, and the monthly routine around it. Writing the pitches, filming the videos, and tracing revenue back to AI answers are separate jobs that start once the queue exists.

Step 1: Write The Questions Buyers Ask At Both Moments

Shortlist questions ask for names: which providers to look at, which suit a team like mine, which have a free plan. Evaluation questions ask for a judgment about a kind of thing: is it good enough for us, does it hold up in a busy month, is it worth the money.

Google Search Central says AI Mode may search related subtopics and sources to build a response. Keep shortlist and evaluation questions separate: Step 4 tests whether they lead to different pages.

Write every question in both shapes and keep the situation identical, so the only thing that changes is what the buyer wants back. These pairs come from our September run:

Use first-call notes for evaluation questions and the lost-deal field in your CRM for the shortlist prospects chose instead.

Write the recorded question first, then its twin in the other shape, and mark the twin as yours. Then ask a salesperson whether a prospect would say it. Twenty questions in each shape is a workable start.

Step 2: Ask Each Question More Than Once And Keep Every Source

Open both engines signed out, or in a browser profile with no search history. Google's AI Mode help page says answers can reflect past searches when personalization is on. OpenAI's search help page says ChatGPT web search works signed out.

In ChatGPT, open a new chat, select View all tools, then Search, and send the question. When the answer arrives, select Sources and copy each link. In Google AI Mode, go to google.com/ai, ask the same question, and copy the links shown with the answer.

Before you move on, check that every block ends in a Sources list or in the words "no sources". A block that ends any other way was cut short in the copying; ask the question again rather than guess what it cited.

Ask every question at least twice on each engine, on different days. When we asked the same question again, the second answer kept on average 40.68% of the sites the first one cited on ChatGPT, and 40.75% on Google AI Mode. A site that shows up once may be gone the next time.

Twenty questions in each shape, two engines, and two asks each make 160 answers. The copying is most of the hours in this audit, which matters later when you decide what an agent should take.

Keep answers with no citations as rows marked "no sources". In September, 7.2% of evaluation answers and 2.4% of shortlist answers cited nothing; deleting those rows would hide that gap.

Two asks a month tell you which sites turned up this month. Which of them stay, and which come and go, shows only when the same questions are asked every day.

Step 3: Let An Assistant Turn The Answers Into Rows, Then Check Its Work

Paste every saved block into an assistant with this prompt, put your company's name where it says <your company>, and save the table it returns as worksheet.csv:

I saved some AI answers by hand and pasted them below. Each answer starts with a line that begins "Q:" and ends where the next "Q:" line starts. Turn them into rows for a citation worksheet.

Write one row for every source link listed under an answer, with these columns:
answer_id | question | shape | engine | date | cited_url | domain | page_title | named

Rules:
1. Copy each URL as pasted, then delete tracking parameters such as utm_source. Never write a URL that is not in the paste.
2. The domain is the site's host without "www." Keep subdomains: help.klaviyo.com stays help.klaviyo.com.
3. If an answer lists no source links, write one row for it with "no sources" in the cited_url column.
4. Take the answer_id, shape, engine and date from the "Q:" line. If one is missing, write "unknown". Do not guess.
5. In the named column, write "yes" only if <your company> appears in the answer text itself, otherwise "no". A source title or a link label does not count.
6. Give me the table only, then one line saying how many answers you read and how many rows you wrote.

The answers:

I tested this in Claude Code on three September answers, with ActiveCampaign as the test name. It returned 18 rows, including one "no sources" row; every URL matched the paste. Here are the first rows with some columns omitted:

It removed the tracking tag ChatGPT adds to its links and kept the parameter that changes the page, HubSpot's ?tier=starter. It marked ActiveCampaign as named in the first answer, the one whose sources never included ActiveCampaign's site.

Three answers prove the prompt runs. They do not prove it never slips on the hundredth, so four checks stay:

• The row count equals the links you pasted, plus one row for each answer with no sources.
• Every URL in the table appears in your paste, apart from the tracking parameters the prompt removes. One that does not is invented; delete the table and run it again.
• One row in ten opens to the page its title describes.
• You read the named column yourself for five answers, because it is the one column that takes reading rather than copying.

Step 4: Build A Separate Working List For Each Kind Of Question

A frequently cited site may add no answers beyond those your list already covers. Choose sites by the answers they add instead.

Ranked by citations alone, the September shortlist sites needed 15 rows to reach four fifths of their answers. The rule below needed eight.

Take the site cited by the most answers and set those answers aside. From the answers left, take the site cited by the most of them, and keep going until the list covers four fifths of the answers that cited anything.

Four fifths keeps the list short enough to open every page; the pool just below it is the next place to look.

Do it once per kind of question, never on the two pooled. This prompt runs it on the worksheet from Step 3, uploaded as a file:

The file worksheet.csv is a citation worksheet: one row per source an AI answer cited, and an answer_id that is the same on every row from one answer. Rows that say "no sources" are answers that cited nothing; leave them out of everything below.

Do this separately for each value in the shape column:
1. Count the answers that cited at least one source. That count is the base.
2. Build the working list. Pick the domain cited by the most answers, then set aside every answer that cites it. Repeat on the answers that are left, each time picking the domain cited by the most of them (break a tie alphabetically), and stop as soon as the answers set aside reach four fifths of the base.
3. Next to each domain, show the running share of the base covered once it is added.

Then, for each shape's list, show the share of the OTHER shape's base that cites at least one domain on it.

Use code for every count. Show the two lists next to each other and mark the domains that appear on both.

On the September worksheet it returned these two lists, each site in the order the rule added it, with the share of that kind of question's answers covered once the site was in:

The list built from shortlist questions reached 58.62% of the evaluation answers, against 80.33% of its own. Put the other way, 41.38% of the evaluation answers cited none of the sites a shortlist-only team would be working on.

The same gap held when we built each engine's lists separately:

These are one category's numbers in one month. In your own sheet, look for the drop from a list's own coverage to its reach into the other kind of question, and how big it is.

Keep subdomains separate: help.klaviyo.com is a help center; klaviyo.com is a homepage. They need different work.

A Site-Type Report Would Call These Two Lists The Same

Sorting the sources by type (vendor sites, roundups, review directories, forums, video) is the obvious shortcut, and on these two lists it hides the split.

A type summary tells you to pitch roundups. It cannot say which roundup, and in this category the roundups split by question: stackfyi.com and zapier.com sat only on the shortlist list, fitsmallbusiness.com and apiscout.dev only on the evaluation list.

Step 5: Open The Cited Pages And Give Each Site One Of Four Verdicts

Open each cited page, check whether it includes your company, and assign one of the four verdicts below. Record what on the page supports your choice.

A page that already includes you is not finished: read what it says about you. If the price, the plan, or a feature is out of date, the task is a correction sent to whoever runs the page; if it is right, note it and move on.

A site whose pages offer different routes gets one row per page.

We opened these cited pages on October 15, 2026, judging them as a hypothetical email platform absent from both lists:

YouTube is Contribute because the route is your own channel: a video of yours on the same question is a new source beside the creators', not a place inside their videos.

Seven of the 15 rows held a task on the site itself, and three of those seven exist only because the evaluation questions were asked: apiscout.dev, fitsmallbusiness.com, and litmus.com. A gap list built from shortlist questions alone would have sent the team to stackfyi.com and zapier.com and never to them.

Two of the seven come with a condition to check first. apiscout.dev lists APIs, so it fits only a platform with a sending API, and litmus.com publishes no submission route, so the pitch starts as a question to its editor.

The six rivals' pages point at work on your own site: check whether you answer the pricing, support, and comparison questions they answer.

The two Watch rows point the same way. The AICPA and CIMA SOC 2 page was cited on the evaluation question about SOC 2, and HubSpot's own security page on the shortlist one.

If buyers in your category ask about a standard you meet, a plain page on your own site that says so is the task.

Google Search Central says AI Mode supporting links must be indexed and eligible for a Search snippet. OpenAI's crawler documentation says opting out of OAI-SearchBot excludes a site from ChatGPT search answers.

Ask whoever runs your website to confirm that neither is blocked.

Three of the seven work rows carry an advertising link (getomnichannel.com, stackfyi.com, and apiscout.dev), and stackfyi.com and apiscout.dev name each other as sister sites. On GetOmnichannel's roundup the link sits in the header, above the ranking it publishes:

Buying advertising is a separate decision with a price on it. The verdict records that the site sells advertising beside its editorial; whether to buy any is the owner's call.

If the lists hold little work in your category, open the pool just below the cutoff as well. The sites tied with the last picks, such as techradar.com, capterra.com, and stackbriefly.com on the shortlist side here, are the next candidates, judged the same way.

Step 6: Put An Owner, A Date, And A Number On Every Task

Whoever owns pipeline signs the worksheet off. Each Contribute or Earn row gets a name and a date. A Competitor's page or Watch row that points at a page your own site lacks becomes an own-site task with an owner; the rest get one line, "check next month".

Write the reason beside every verdict, in a sentence. It is what lets someone else, or an agent, apply the same judgment next month.

Prioritize rows that reach the most answers for the least effort. In this hypothetical queue, each task is due before next month's audit, except the robots.txt check, due this week:

Then decide what you will count, and what each count can and cannot tell you:

In Gartner's survey of 645 B2B buyers, 45% used generative AI during a recent purchase, mostly for vendor information, and 69% prefer to validate AI-generated insights with a sales rep.

So ask on that call what the buyer read before they called, and write it down.

None of these counts proves the work made money. A citation and a sale sit far apart, and the question on the first call is the nearest link between them; it rests on what a buyer remembers.

Have An Agent Draft Each Month's Queue For The Owner To Approve

An agent can repeat the extraction, arithmetic, and comparison each month. The owner still judges new pages and approves the work.

Use the daily answer and citation records from the GEO Agent, or the answers your team saved by hand, as the inputs to this monthly brief.

The agent runs six steps a month, written here so you can brief whoever builds it:

1. 01

   Read the month's records

   The month's answers and their sources, from the GEO Agent's records or the blocks your team saved; last month's worksheet with every verdict and its reason; the month's first-call notes and lost-deal reasons from the CRM; and your own website.

2. 02

   Run on the first working day of the month

   One run a month, on the same questions, so this month's lists compare with last month's.

3. 03

   Rebuild both lists by the same rule

   Four fifths of each kind of question's answers, with the ties shown, then a comparison with last month: sites that joined, sites that left, and sites whose cited page changed.

4. 04

   Carry every verdict that still holds

   A site keeps last month's verdict while the engines cite the same page, and every carried page goes back to the owner to reopen once a quarter, because a page can change at the same address. A new site, or a known site cited on a new page, goes to the owner marked "to open", with the page links and the answers that cited them.

5. 05

   Draft the own-site gaps and the pitches

   For each rival's page or Watch row, look for a page on your own site that answers the same question and note what is missing. For each Earn row the owner has approved, draft a pitch against the page's own stated method.

6. 06

   Hand the owner one queue

   Carried rows, rows to open, rows that left, and drafts, in one sheet. The owner rules on every new row and every draft.

Count it the way you counted the first audit. The owner's hours on the monthly review sit beside the hours the hand-run audit took, and the rows the owner corrects are the agent's error count.

How Qvery Keeps Your Citation Queue Moving

We begin your citation audit with the GEO Agent's records. It prepares questions using what you sell, the industry you serve, and the competitors you face.

It runs those questions every day. ChatGPT and Google AI Mode are its two destinations, with countries chosen by you. It retains responses and their complete citation lists.

You can inspect sources by country or topic and compare sentiment with rivals. Visibility and competitor share of voice sit alongside that evidence. We go through the findings together, then handle corrections and content work on the pages the audit identifies.

When rebuilding the queue takes too much time, we learn how your owner reaches each verdict and create a custom agent for that monthly routine. It runs on autopilot in your existing systems, preparing changed rows and pitch drafts for approval.

Your queue owner learns to review the agent's output as we get the first version running within two weeks. We tally audit preparation hours saved and row errors avoided each week, using the manual audit as the comparison.

Your worksheet gives us the decisions the agent must learn to carry forward.

Compare The Second Audit With The First

Run the second audit a month after the first and lay the two worksheets next to each other. If the top four rows of each list held and the hours stayed small, keep doing it by hand.

If the rows churned and the copying grew, an agent's build and review is worth pricing against those hours.

Either way, keep the first month's verdict column with its reasons. It is the one part of the audit nobody can rebuild from the answers, and it is the rulebook an agent drafts against in month two.

The same decision about repeated work applies to the call notes your salespeople enter into the CRM.