How to Run a Content Gap Analysis for AI Search

A content gap analysis, as SEO tools run it, lines up the search terms your site ranks for against a competitor's and hands back the topics they cover and you don't. It was built for a results page with ten blue links.

Before you approve the next batch of articles from that list, know what it cannot see.

Your buyers now ask an assistant, and the assistant answers from a shelf of pages it cites. Which pages sit on that shelf depends on the question.

So the gap worth finding is a missing kind of page for a kind of question. A site stocked only with product pages has nothing on the shelf where learning questions get answered, however many topics its blog lists.

By hand, you label cited pages and compare them with your sitemap. An assistant handles the labeling and matching; a person decides what gets written.

We tested the method on learning and buying questions about security tools for small companies.

Use the two tested prompts below to turn your buyers' questions into a gap table with an owner for each task. After the steps, I explain how we keep that table current at Qvery.

One limit before the steps: it finds the kinds of page your site is missing for each question. It cannot promise that a new page gets cited. OpenAI's own help page says ChatGPT ranks search results on several factors and does not promise placement.

What You'll Need

This finds the pages your own site is missing and who writes them. Getting onto other companies' pages is a different job, a citation gap analysis, and writing the pages starts once the table exists.

Step 1: Write Every Buyer Question Twice, Once to Learn and Once to Buy

Learning questions ask how something works or whether the buyer needs it. Buying questions ask which one, from whom, at what price. The same buyer asks both, often weeks apart, and the answers to the two are built from different shelves.

Write each question in both shapes and keep the subject and the buyer's situation identical, so the only thing that changes is what the buyer wants back. These pairs come from the September questions:

Take them from your own records, never from a guess about what people type.

The learning questions are in your first-call notes and in the emails prospects send before they ask for a quote; the buying questions are in your lost-deal reasons and in how prospects describe the shortlist they picked from. We wrote up how to turn call notes into questions separately.

Every industry has both shapes of question. These are illustrations of the pairs to listen for, not questions we measured:

In the security questions, the short list of sites that covered four fifths of the learning answers reached only 56.78% of the buying answers that cited a source. In email marketing software the hole ran the other way: 41.38% of the evaluation answers cited none of the sites a shortlist-only team would work on.

Twenty questions in each shape is a workable start.

Step 2: Ask Both Engines and Keep Every Cited Page With Its Title

Open ChatGPT signed out, which OpenAI's search help page says supports web search. Open Google AI Mode at google.com/ai in a browser profile with no history: Google's personalization help page says previous searches can affect AI answers.

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 with the title shown beside it. In Google AI Mode, ask the same question and copy the links shown with the answer.

Save the title with every link. Step 4 decides what kind of page each link is from the link and its title, so a link saved without its title is a row the assistant will have to mark "check".

Ask every question at least twice on each engine, on different days.

Google's help page says AI Mode works by "dividing your question into subtopics and searching for each one simultaneously", and it advises people to "Ask multiple versions of your question". When we re-asked the email marketing questions, the second ChatGPT answer kept on average 40.68% of the sites the first one cited.

Keep answers that cite nothing as rows marked "no sources"; a deleted row hides a trend. In our research, 86.4% of learning answers cited a source, against 94.4% of buying answers.

Two asks a month tell you what turned up this month. Which pages stay on the shelf for a question, and which come and go, shows only when the same questions are asked every day.

Step 3: Export the Pages Your Own Site Already Has

Open yourdomain.com/sitemap.xml in a browser. The sitemap protocol recommends placing it at the root of the site; if it is not there, open yourdomain.com/robots.txt and look for a line that starts with Sitemap:. If the file lists other sitemaps instead of pages, it is an index: open each one.

Copy each URL into a sheet with three columns: url, title, first_paragraph. Get the titles and opening paragraphs from your content system. The first paragraph matters because Step 5 judges what each page says up front.

Keep product, service, pricing, help, guide and blog pages. Leave out tag pages, legal pages and logins.

Check the row count with whoever runs the website before you go on. A sitemap that lists far fewer pages than they expect is missing some, and a page missing from your list will show up later as a gap you do not have.

If there is no sitemap at all, Search Console's Page indexing report lists the pages Google knows, though its list of example URLs "is limited to 1,000 items".

To show the next steps on something concrete, I made up a company: Harbor Backup, a cloud backup provider that sells to small businesses, with eight pages. The company and its pages are invented; the questions and answers it is measured against are the real September ones.

Step 4: Have an Assistant Sort Every Cited Page by Kind

Paste your saved answers into the assistant with this prompt and save the table it returns as sorted.csv:

Below are AI answers I saved by hand. Each one starts with a line that begins "Q:" and ends where the next "Q:" line starts. Under each answer is the word Sources: and then one cited link per line, sometimes followed by " | " and the page title.

Sort every cited page by what kind of page it is. Write one row for every link under Sources:, with these columns:
answer_id | question | kind | engine | cited_url | site | page_kind

Use exactly one of these for page_kind:
- public body: a government agency, regulator, or standards body
- company guide: a company's own article, guide, explainer, glossary, or blog post
- company sales page: a company's product, service, pricing, or home page
- roundup: a "best of", comparison, or alternatives page, whoever publishes it
- review site: a site of user reviews or ratings
- forum: a forum, community, or Q&A thread
- video: a video page
- reference: an encyclopedia, a trade publication, or a news article
- check: you cannot tell from the link and the title

Rules:
1. Copy each URL exactly as pasted, then delete tracking parameters such as utm_source. Never write a URL that is not in the paste.
2. The site is the host without "www."
3. Decide page_kind from the link and the title only. If they do not settle it, write check. Do not guess.
4. If an answer has "no sources", write one row with "no sources" in cited_url and leave site and page_kind empty.
5. Return only the table, as CSV with a header row.

We ran it in Claude Code on forty saved answers to eight backup questions from the September run. It took just over a minute and wrote one row per link, plus one row for each answer with no sources; every URL in the table matched a link in the paste. The first rows:

It marked five rows "check" rather than guessing. Against our own classification of the same links it agreed on 87.97% of rows, and most of the disagreements were arguable: it called Microsoft's 365 backup page a sales page and a site's Backblaze-versus-IDrive page a roundup, both fair.

The rows it wrote do not prove the next hundred will be right, so four checks stay:

• The row count equals the links you pasted plus one row per answer with no sources.
• Every URL in the table, apart from the tracking parameters the prompt removes, appears in your paste. One that does not is invented: delete the table and run it again.
• You open every row marked check and label it yourself.
• You open ten other rows at random and confirm the label matches the page.

Learning Questions Are Answered by Guides, Buying Questions by Product Pages

Across the September questions, the sorted table shows which kinds of source each side cited:

The row that decides the content gap is the companies' own sites, and which of their pages got cited:

These are one category's answers in one month. In your category the split is the thing to measure, not to assume.

For "Business backup basics", the answers cited Huntress's post on the 3-2-1 backup rule, Veeam's blog, Cloudian's backup guide and HPE's glossary entry beside CISA.

For "Recommend backup software for a five-person company", they cited Backblaze, Acronis, Veeam, IDrive, and CrashPlan, almost all on product and pricing pages.

The single page cited in the most learning answers was CISA's #StopRansomware Guide, in 9.26% of the learning answers that cited a source:

A company page will not replace CISA as the source on ransomware. What it can add is how the advice applies to its own buyers.

CISA's #StopRansomware Guide tells organizations to "Maintain offline, encrypted backups of critical data" and test them regularly. Your page can explain how to apply that advice to a small office and test a restore on your product.

Three kinds of row look like gaps and are not:

1. Roundups and review sites

   The shelf belongs to an editor or a reviewer. Hand the row to the citation gap queue.

2. Forums and videos

   People are answering each other. Taking part is a separate decision with its own rules.

3. Public bodies on a question you have no stake in

   The agency is the right answer. Leave it, and note why.

"What happens after a phishing breach at work?" was answered from the NCSC's incident pages and the UK data regulator's breach guide. Unless your company handles incidents, there is no page of yours to write there.

Step 5: Lay the Cited Pages Against Yours and Give Each Question a Verdict

Upload sorted.csv and your page list as pages.csv to the assistant with this prompt:

I am attaching two files.

sorted.csv lists every page that AI answers cited for our buyers' questions, one row per cited link, with the question, its kind (learning or buying), the engine, the cited URL, the site, and a page_kind.

pages.csv lists every page on our own website: its URL, its title, and its first paragraph.

Our site is the site in pages.csv. For each question in sorted.csv, write one row with these columns:
question | kind | answers | cited_page_kinds | ours_cited | closest_page | verdict | reason

How to fill them:
1. answers: how many answers to the question cited at least one link.
2. cited_page_kinds: each page_kind that appears, with how many of those answers it appears in, most first. Example: "public body 4, company guide 3".
3. ours_cited: yes if any cited link is on our site, otherwise no.
4. closest_page: the one URL from pages.csv whose title or first paragraph answers the question most directly, or "none". Judge only from the title and first paragraph. Do not assume a page says anything it does not say there.
5. verdict, exactly one of:
   - Already cited: ours_cited is yes.
   - Keep and re-ask: we have a page of the kind the answers cite most, and its first paragraph already answers the question.
   - Rewrite the opening: we have a page of the kind the answers cite most for this question, but its first paragraph does not answer the question.
   - Write the missing page: the answers cite most a kind of page we do not have for this question.
   - Apply the public guidance: public bodies are the most cited kind; we have no page that explains how that guidance applies to our buyers.
   - Not a page job: roundups, review sites, forums, or videos are the most cited kind.
   - No sources: no answer to the question cited a link.
   If two kinds tie for most cited, say so in the reason and judge by the one that comes first in this order: public body, company guide, company sales page, roundup, review site, forum, video, reference.
6. reason: one sentence, quoting the words from our title or first paragraph that you relied on, or saying that no page of ours mentions the subject.

Never invent a page, a URL, or a count. Return only the table, as CSV with a header row.

Each question gets one of seven verdicts. The table names possible gaps; a person confirms each one against the full page before anyone writes:

We ran it on the sorted backup answers and Harbor Backup's invented page list. It took under forty seconds, and every count it wrote matched the sorted rows when we checked them:

Harbor's pricing page opens with "Per-device plans with unlimited storage, billed monthly or yearly" but gives no price. The owner decides whether to publish one; if so, put it up front to answer the thirty-dollar question.

"Business backup software recommendations" tied between sales pages and roundups, and the tie rule sent it to the sales-page verdict; a person might add it to the citation queue as well.

The home page "never mentions five people", as the assistant's own reason says, so a person may rule that row a rewrite instead. The closest page column is a judgment from a title and a paragraph: open each page before anyone writes a brief.

Step 6: Put an Owner, a Date, and a Re-Ask on Every Row

Whoever owns pipeline signs the table off. Each row that leads to work gets a name, a date, and a re-ask date: the day the same questions go back to both engines to see whether your page is now on the shelf.

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

For Harbor, the finished shortlist would read like this, owners named by role. The five-person question folds into the backup basics guide, and the two rows marked keep and re-ask need only the re-ask:

All eight questions go back to both engines in week 8, the re-ask date for every row.

Before re-asking, have the web lead check that the new pages appear in your sitemap and allow OAI-Searchbot to crawl them. OpenAI's search help page names crawler access as a condition for inclusion.

Decide with the owner what to count after each round:

None of these proves the pages made money. A citation and a sale sit far apart, and the nearest link between them is a question on the first sales call: what did you read before you called us?

Have an Agent Redraw the Gap Table When Your Pages or the Answers Change

The second round repeats the same sort and verdict rules on new inputs. That is the job an agent can take over.

After two manual rounds, keep your reasons and hand this brief to whoever builds the agent:

1. 01

   Reads

   The month's answers with their sources, from the GEO Agent's daily records or the blocks your team saved; your sitemap; last month's gap table with every verdict, owner and reason.

2. 02

   Runs

   On the first working day of each month, and again whenever the sitemap gains a page.

3. 03

   Applies

   The sort rules of Step 4 and the verdict rules of Step 5, unchanged, so this month's rows compare with last month's.

4. 04

   Produces

   A draft gap table with the changes marked: new questions, rows closed because your page is now cited, rows reopened, verdicts that moved and why.

5. 05

   Is approved by

   The owner, row by row, before any work is assigned.

6. 06

   Is counted in

   The owner's review hours against the hours a hand-run round took, and the rows the owner corrects, which are the agent's error count.

If your questions and site rarely change, a calendar reminder may be enough. When both change monthly, compare the cost of a build with the hours each manual round takes.

How Qvery Keeps Your Content Gap Table Current

Your gap table gives us the decisions to learn from. We review how you approved its rows, then build a custom agent around your content system and working sheet. It redraws the table on autopilot when the monthly review comes due or your sitemap changes; your owner approves the proposed work.

Our GEO Agent supplies the answers for that review. It uses your brand, the industry you serve, and your competitors to compose buyer questions. Each day it collects responses from Google AI Mode and ChatGPT, with citations attached to each answer and the country chosen for that run.

You can inspect the cited sources by topic or country, alongside your share of voice, your visibility relative to competitors, and sentiment toward each company. We read that evidence with you and do the work on the pages it points to.

Within two weeks, your content lead has a working table agent and our training on checking changed verdicts. We record its effect weekly: hours saved sorting and matching, and errors it prevented, including mistaken briefs for pages your site already has.

Bring the sheet with your reasons still attached so we can learn the judgments behind the rows.

Start Where Other Companies' Guides Already Get Cited

If the table can carry one row this month, pick the learning question whose answers already cite other companies' guides. That shelf has already shown it takes a company's page.

For Harbor, write the backup basics guide, apply CISA's advice to a small office, and link the agency. Put that purpose in the opening paragraph and re-ask in eight weeks.