DeepSmith

Nov 26 · AEO & AI Visibility

19 min read

Which Content Formats Win AI Citations? How to Find Out From Your Own Data

Avinash Saurabh
Avinash Saurabh · CO-Founder & CEO
Monochrome illustration of layered list, step and side-by-side page cards linked to a small bar chart, under the headline Which formats get cited.

If you want to know which content formats ai citations favor in your category, the most reliable answer comes from your own tracking data, not from an industry study. This guide shows you how to take the pages that AI answers cite for your buyer questions, tag each one by format (listicle, how-to guide, comparison, and so on), and compare the formats fairly. By the end you will have a small table that says which format your category rewards, for which kinds of questions, and how sure you can be.

You need a set of tracked prompts with the URLs the AI engines cited, and a spreadsheet. Plan on a few hours for a first pass, mostly spent on the tagging.

Why published studies do not settle this

Several companies have published studies on what content gets cited by AI, and they do not agree with each other, which is a useful hint about why you should measure your own.

A Wix Studio and Peec AI analysis looked at 75,000 AI answers and more than a million citations from ChatGPT, Google AI Mode and Perplexity. The Wix analysis put listicles first overall at 21.88% of citations, with articles at 16.68% and how-to guides at 6.21%. But the leader changed with the type of question. For informational prompts, articles led at 45.48%. For commercial prompts, listicles led at 40.86%. For transactional and local prompts, product pages came out on top. The study also said that the type of question predicted the format mix better than the industry or the AI model did.

An Otterly AI study of about 618,000 citations reported listicles at 66.6% of classified citations. That sounds like a different world from 21.88%, and the gap mostly comes from the prompts. Ranking questions made up about three quarters of the prompts in that study, and listicles were cited far more often for ranking questions than for open-ended ones. The article also does not clearly say whether its percentages are shares of unique URLs or of all citation occurrences.

A Seer Interactive analysis tracked a consistent prompt set for about three and a half months and used URLs containing "best" or "top" as a stand-in for listicles. It found ChatGPT listicle citations fell about 30% from December to January, while total citations also fell 22.7%. The write-up also noted that ChatGPT and Google AI Overviews moved in opposite directions on the same prompts. That is one month-to-month comparison with a partial final month, so it does not show that listicles are losing in general.

None of these studies is wrong, but each one used its own prompts, engines, dates, and definitions of "listicle." So they tell you that format depends on the question and the engine. They do not tell you what your category rewards. Treat the numbers as background, never as targets for your own pages.

The rest of this guide covers the eight steps for finding your own answer.

Step 1. Fix the question set you want to answer

Start with buyer questions that belong to one topic or category. Write down the exact prompt text for each one, along with its intent (informational, commercial, transactional, or navigational), its topic, and its funnel stage. Use the real questions your buyers ask, not only the ones that happen to include "best" or "top."

You are done when you have a named group of prompts, each recorded and labeled, and you can say in one sentence which buyer task the group represents. These intent labels are working buckets for your own data, so define them once and stick with them.

Common mistake. Pooling broad "how do I" questions with "which vendor should I buy" questions and treating the combined format mix as one result. The format distribution changes with intent, so a mixed pool hides the difference.

If you track prompts in DeepSmith AI Visibility, this is where it helps. You define the buyer prompts yourself, or use Discover Prompts to generate a starter set from your product and persona context. The platform checks them on a schedule and keeps the full answer history, which gives you a repeatable group. It does not label formats for you, so the classification in the later steps stays your job. If you are still choosing a tracker, there are roundups written for content teams, for marketers and for B2B teams.

Step 2. Hold the measurement cohort steady

Before you count anything in your content formats ai citations study, decide which AI engines, which prompts, which date window, and which market you are including. Then keep those the same from one period to the next. If your tool lets you filter by date or platform, use that to build matched slices. Write down any change you make to prompt wording or tracking setup.

You are done when you can describe exactly which responses sit in each analysis and someone else could rebuild the same selection next quarter.

Common mistake. Comparing last month with this month after adding prompts or engines. A rise in listicle citations might only reflect a change in your question mix, and not a change in how engines treat listicles.

Repeated collection matters here because citations move around. An Ahrefs analysis of 37,500 citations found that around 70% of new citations were gone the next day, and "best X" questions were among the least stable. That is a finding from one study, not a rate you can apply to every engine. What it suggests for you is to look at a trend across several collection windows and to avoid reading much into a single answer.

Step 3. Collect the cited URLs and their response context

Build a working table with one row per citation occurrence. Keep at least these columns: the prompt text or ID, the engine, the response date or run, the cited URL, the cited domain, and whether the URL is yours, a competitor's, or someone else's. If you can, keep a link or ID for the underlying answer so you can check any row that looks strange.

You are done when every citation in scope ties back to a prompt, an engine, a date, and a URL, and you can pull out the subset that belongs to your own site.

In DeepSmith, the Pages view shows which of your pages AI cites, each page's share of your citations, and the prompts driving them. Competitor citations show who wins for your prompts and on which exact pages. If you would rather pull the data into your own tools, the read-only API and the MCP server both expose cited pages and prompt answers. The MCP tools include list_cited_pages, get_cited_page, get_competitor_pages, get_cited_sources and get_prompt_answers, and get_prompt_answers returns up to 200 answers for a prompt, with the answer text and citations. The MCP server has no bulk export, so for a large set you will be copying or scripting in batches. Neither the API nor the MCP server assigns editorial format labels.

If your question is wider, such as which sites AI engines lean on across a whole industry, a separate walk-through on finding the sources AI engines cite covers that. Here the focus is narrower, on what kind of page gets cited. For a tool-by-tool view of the wider question, see the comparison of tools that show which sources AI engines cite.

Common mistake. Mixing your own pages with competitor and third-party pages in one denominator and then calling the result "our content formats." Analyze your owned pages on their own. Competitor and third-party citations are useful for comparison, but they are a separate group of pages, and a citation source audit is the better place to work through them by source type.

Step 4. Clean and classify each cited page by editorial format

First, clean the URLs. Strip query strings and fragments, follow redirects, and merge trailing-slash variants, so that one page does not show up as three. Then open each page and give it one primary format label. A controlled list keeps this consistent. The table below is a practical starting set for what you might use.

Primary formatWorking definition for this analysis
Listicle / roundupA ranked or enumerated selection of several options, such as "best tools" or "10 examples."
How-to guideA step-by-step page that explains how to complete a task.
ComparisonA direct side-by-side of two or more choices.
Article / explainerEditorial or informational coverage that is not mainly a listicle, task guide, or head-to-head comparison.
Product pageA page about one product or service, usually with features, specs, or purchase details.
Category pageA collection or index page that groups products, services, or other pages.
Discussion / forumA user thread, question-and-answer page, or community post.
News / researchA news report, original research, or a study.
Other / unclearAnything you cannot assign with confidence.

This is a working taxonomy for your own analysis, and not a claim that every study defines these the same way. Some studies also split out homepages, profiles, alternatives pages, documentation, video, and social posts. Add a category only if it matters for your category and you can apply it the same way every time. Keep the "Other / unclear" bucket visible and do not force a label onto a page that does not fit.

Store the format label in its own column, separate from the other facts about the page such as topic, intent, funnel stage, publisher type, and whether you own it. For instance, a SaaS comparison article could be tagged comparison for format, commercial for intent, consideration for funnel stage, and owned for ownership. Keeping these apart lets you slice the data later without making the format list unwieldy.

You are done when every canonical URL has exactly one primary label, an "other" option exists, and a second reviewer could understand the rule you used. A good test is to have a colleague tag twenty pages on their own and compare against yours.

Common mistake. Deciding the format from the slug or title alone. A page called "guide" may really be a product comparison, and a title with "best" in it is not always a listicle. Look at the page, and settle the unclear ones with the same rule each time. The Seer study's URL-text shortcut is a good example of why this matters, since it counts any URL with "best" or "top" as a listicle.

Step 5. Build the citation format analysis with explicit denominators

This is the center of a citation format analysis. Calculate at least three measures, and write the denominator next to each result.

  1. Citation share by format. Citation occurrences for a format divided by all in-scope citation occurrences. It answers what share of the citations in this sample went to that format.
  2. Citation events per eligible page. Citation occurrences for a format divided by the number of eligible pages of that format in your inventory. This helps separate how well a format does from how many pages of it exist. Read it as a description of what happened, and not as a promise for a new page.
  3. Cited-page coverage. Distinct cited URLs in a format divided by all eligible URLs in that format. It answers what share of your pages in that format were cited at least once.

A fourth measure is worth adding when you can. Run prevalence is the number of eligible prompt-engine runs in which the format appeared at least once, divided by all eligible prompt-engine runs. It shows whether a format's citations are spread across many answers or packed into a few.

Define the unit of counting before you start. A workable rule is one occurrence per canonical URL per response, so a URL mentioned twice in one answer counts once. Keep unique URL counts and citation occurrence counts separate, since a page cited fifty times is still one page.

Here is a worked example. The numbers below are invented to show the arithmetic. They are not measured results and not a customer outcome. Suppose a fixed cohort produced 600 owned-site citation events across 1,000 eligible prompt-engine runs, and the site has 400 eligible pages.

FormatCitation eventsEligible pagesDistinct pages citedRuns with format citedCitation shareEvents per eligible pageCited-page coverageRun prevalence
Listicle240803021040%3.037.5%21%
How-to guide1802006016030%0.930%16%
Comparison120201210020%6.060%10%
Other formats60100205510%0.620%5.5%

In this made-up case, listicles have the largest citation share, at 40%. Comparisons have the highest events per eligible page and the best cited-page coverage, because there are only 20 comparison pages and they are cited a lot relative to their number. How-to guides are the biggest part of the inventory and get fewer citations per page. Neither reading means comparisons would win for a different prompt set or category. The point of the example is that raw volume, coverage, and yield per page can each tell a different story, and you should report all of them.

When a format has very few eligible pages or very few observations, mark the result as low-volume. A format with three pages and two citations can rank first or last by chance.

You are done when your output table shows, for each format, the citation occurrences, unique cited URLs, eligible inventory pages, citation share, events per eligible page, cited-page coverage, and ideally run prevalence, with every denominator written down.

Common mistake. Calling the format with the highest raw citation count the best-performing format. Raw totals favor the formats you have the most pages in, or the ones your prompts happen to touch most. Compare volume, coverage, and yield per page together.

Step 6. Pivot the data by format, engine, and intent

A spreadsheet pivot table does this part well. Put the primary format in the rows, summarize citation occurrences as the value, and put engine and intent in the columns or filters. If your source table has them, add unique cited URLs and inventory page counts as separate values. Filter to your owned pages when the question is which of your formats gets cited, and build separate tables for competitor and third-party citations if those matter to you.

In Google Sheets, select the source range with a header on every column, then choose Insert and then Pivot table. In the pivot editor, add your fields under Rows, Columns, Values, and Filters. Google's help pages on how to create and use pivot tables walk through the editor. Keep in mind that the pivot only aggregates the records you tagged. It does not classify any pages for you.

You are done when you can answer two questions for the whole cohort and for matched engine and intent slices: which format has the most citation events, and which has the strongest coverage or yield within the pages available.

Pro tip. Make the engine-by-format table before the grand total. The studies above showed the engines disagreeing, with Perplexity elevating discussions where ChatGPT and Google AI favored articles in second place in the Wix data. If your own pivot shows a format that wins on one engine and is absent on another, a single pooled number would have hidden that.

Common mistake. Averaging unlike engines and intents into one total and calling it your category's stable preference. A pooled number can conceal a format that wins in only one place.

Step 7. Interpret the leaders without overclaiming

Look at whether the leading format stays near the top across repeated periods and across matched slices. Note where the results differ by engine or by intent, and whether the format has enough pages and observations for the comparison to mean anything. Treat the leader as a hypothesis for your content plan, and avoid turning every topic into that format because of one table.

You are done when your readout states the winner for the specified cohort, the metric you used, which segments it covers, the measurement period, the sample size, and the main uncertainty. For example, "Among owned pages, comparisons had the highest citations per page for commercial prompts on two engines over three months, based on 20 pages" is a statement someone else can check and challenge.

Common mistake. Saying "AI cites listicles" without saying which prompts, engines, dates, market, or denominator. Studies that report very large listicle shares often have prompt sets weighted toward ranking questions, and the same would be true of your own data if most of your tracked prompts say "best."

Content Map in DeepSmith can add some context here. It shows topic coverage and funnel-stage distribution for your site and your competitors, so you can tell whether a result reflects a narrow topic or a wider gap in what you publish. It is a way to understand coverage, and it does not replace the format-tagged analysis unless you supply the tags and calculations yourself.

If what you see is that your category rewards a format you have few pages in, the roundups of tools that find AI citation opportunities can help you spot which prompts are worth targeting with it.

Step 8. Turn the result into a controlled content test

Pick a topic you have not covered, or a page that already earns citations, and create or update it in the format your data supports. Before you publish, write down the target prompt group, the engine, the baseline, the publish or update date, and the main success metric. Then re-check the same prompt and engine group over the following collection windows, comparing like for like and noting anything else that changed in the meantime.

You are done when the test has a baseline from before the change, a defined format label, a fixed measurement group, and a review date on the calendar.

Common mistake. Claiming that the format change caused a rise in citations when the before and after periods also differed in prompts, engines, or how often you collected. This kind of analysis finds associations and opportunities. It does not isolate cause, so word your conclusions to match.

DeepSmith Opportunity Agents can help with picking what to test. They read your AI visibility data or your Content Map and return content ideas, and each idea carries the data point that justifies it. That gives you a backlog you can defend. An agent does not prove that a given format will earn citations, so the test in this step is still what tells you whether the idea worked.

What to do next

Choose one format hypothesis for one topic, record the baseline, and come back to the same prompt and engine group after the page is published or refreshed. If your tracking is not yet in place, the roundup of AI citation tracking tools compares options for the page-level data this analysis depends on, and the broader AI visibility tools list covers the wider category. Teams that want cited pages, prompts, and content production in one place can start a 7-day DeepSmith free trial and inspect the pages and prompts behind their citations.

Frequently asked questions

Which content formats win the most AI citations?

There is no universal winner. Listicles can dominate samples that are weighted toward ranking prompts, while articles, how-to guides, product pages, category pages, comparisons, or discussions may lead for other intents and engines. Use your own fixed prompt group and report the denominator next to every result.

How do I find what content gets cited by AI in my category?

Collect the cited URLs and answer context for a steady set of category-specific prompts and engines. Classify each URL by primary editorial format, then compare citations, unique cited pages, and available inventory by format. Split the analysis by intent and engine so a single pooled number does not hide the differences.

Should I count each URL once or every time it is cited?

Keep both measures. Count each citation occurrence within a defined response for citation volume, and count each canonical URL once for unique-page coverage. State how you handle a URL that appears twice in one response.

How often should I repeat the analysis?

The sources reviewed do not set a universal schedule. Repeat it on a regular schedule that fits your monitoring program, with the same prompts and engines so the periods can be compared. Avoid treating one answer or one month as a stable result.