AI Citation Tracking: How to Track LLM Sources That Change Every Day

AI citation tracking is the practice of recording which web pages AI search engines such as ChatGPT, Perplexity, and Google's AI Overviews cite as sources in their answers, how often a brand's own pages are among them, and how that changes over time.

The obvious check is to ask ChatGPT about the category and look for the website among the links, yet the links change from one answer to the next. Across five stored ChatGPT answers to one buying question about AI tracking tools in September 2026, Atomic recorded 13 different cited pages, and no page appeared more than twice.

For a SaaS, fintech, or AI company, one answer can make a missing citation look permanent or hide a gap that recurs every week. This guide shows how to track AI citations as rates on a fixed question set: the six metrics worth measuring, a seven-step setup, how to tell a real change from a daily swing, and how to turn recurring source gaps into content and outreach work.

Key takeaways

• Report each citation rate with its answer count, because two rates mean something side by side only when both show how many answers sit behind them.
• Judge a change by its own margin, which for independent windows of equal size at similar rates is about 1.4 times either rate's margin.
• Count citations and mentions separately, because engines link pages without naming the brand and name brands without a link.
• Work first on the recurring third-party pages that leave the brand out, once each one checks out as relevant to the product.
• Write the question set in customer wording and freeze it, because a rate describes only the questions behind it.
• Compare every change with a control, because an engine update can move citation counts across the whole web, edited pages or not.

What is AI citation tracking?

An AI answer that searches the web can name brands in its text and link web pages as its sources, on average three or four domains in a ChatGPT search answer. A citation is one of those source links. A mention is the brand's name in the answer, linked or not.

AI citation tracking records both across many answers, which is how teams working on generative engine optimization (GEO), or answer engine optimization (AEO), measure results. Tools often sell it as prompt tracking.

A tracker also records how each answer frames the brand. Suppose one answer calls Mixpanel the leading option for product teams and another calls it one option among many. Both count as a mention, yet only the first recommends it.

In this guide's worked example, Vendor A, a B2B SaaS analytics company, finds a ChatGPT answer that cites its pricing page. The same answer calls the product "strong but lacking a warehouse sync on its starter plan," a limit Vendor A removed in August. That one record holds a citation, a mention, a negative framing, and a factual error.

In one ChatGPT answer from October 2026, the text recommends Omnius, the company Atomic began inside. A source chip beside the name links Omnius's case-studies page, which also heads the Sources panel, so the same brand gets a mention and a citation in one place.

The two often happen apart, though. In a Semrush study of 3,981 domain appearances in AI answers, 61.7% linked the brand's page without naming the brand. The study calls these ghost citations. Another 25.1% named the brand with no link, so a report that adds the two together cannot show which one moved.

Why AI citation tracking matters for SEO and GEO

Rank tracking compares a keyword's position from day to day, while AI citation tracking reads the sources inside answers that are written fresh each time. Larger studies show that the churn in Atomic's five answers is common.

Glen Allsopp's study checked more than 1,300 prompts daily for four weeks. In that time, the typical software prompt never returned the same cited links on any two days.

SISTRIX followed 82,619 prompts in six countries for 17 weeks and measured churn-in, the share of a week's cited domains that the previous week had not cited. In Google AI Mode, churn-in ran at 54% to 59% a week in every country.

ChatGPT search varied more, from 74% in Germany to 42% in France. A single check records only the sources in that one answer, and a trend needs the same questions asked again and again.

Rank reports also miss most cited pages. Ahrefs found that 37.1% of the URLs cited in Google AI Overviews in its early-2026 data also ranked in the top 10 organic results for the same query. A rank tracker limited to Google sees nothing from ChatGPT or Perplexity either.

Traffic reports miss them for a different reason, because an answer read without a click leaves no visit to record. The answers themselves are the most direct record of AI search visibility, and AI citation tracking samples them on a schedule.

Why the sources change from one answer to the next

Google says AI Overviews and AI Mode may turn one question into several related searches, which it calls query fan-out. OpenAI says ChatGPT search rewrites a question into targeted queries and may send more specific ones after reviewing the first results.

The diagram groups these documented behaviours into four conceptual stages. A new answer can take a different path through any of them.

The narrower searches can carry the asker's context, such as an approximate location. So a page can be cited for a sub-question even when it does not rank for the original one.

OpenAI's developer documentation notes that "the number of sources is often greater than the number of citations." A citation rate counts only the sources an answer displays, while some tools record retrieved pages separately.

The AI platforms also differ from each other. In SISTRIX's data, AI Overviews and AI Mode shared only 17% of their cited domains for the same prompt. Developer APIs also take settings the consumer apps lack, so each rate holds for one engine and one way of collecting answers.

Signed-in answers add a variable. ChatGPT can shape an answer with past chats, saved memories, custom instructions, and files, and its Sources "may not show every factor that shaped a response." A panel without account history therefore uses logged-out chats and records that state, and its results describe those conditions.

What each AI engine shows, and what first-party data sees

The website's own analytics see only part of each engine. Google counts AI Overviews and AI Mode inside Search Console's normal Web search numbers. GA4's AI Assistant channel leaves both out, so under the default rules visits from Google's AI features arrive as Organic Search.

Bing's AI Performance report, in preview since February 2026, gives citation counts by page for Copilot and Bing's AI summaries. It also shows the trend and the grounding queries, the searches Microsoft's AI ran to find the pages it cited.

First-party reporting on AI search expanded in 2026:

• 13 May 2026: GA4 added a default AI Assistant channel for visits referred by assistants such as ChatGPT and Gemini.
• 3 June 2026: Google launched generative AI reports in Search Console, with impressions and no clicks; they reached all websites on 31 August.
• 16 June 2026: Bing Webmaster Tools added Citation Share, a website's share of the citations shown for one grounding query.

None of these reports shows the answer text. Search Console and Microsoft's reports cover their own surfaces, and GA4 records only the visits it can attribute. A question set tracked across engines still needs saved answers.

Six metrics to measure AI citations

Answers vary between runs, so the numbers that hold up are rates and shares over a fixed window, on the same questions, engines, and collection conditions. Each is read per engine and per question cluster, a group of questions with one intent such as pricing:

1. Citation rate: how often AI engines cite a website

Citation rate is the share of tracked answers that cite a given website or page. In the worked example, Vendor A tracks 20 pricing questions in ChatGPT once a day for 28 days. That gives 560 answers, and 123 of them cite a Vendor A page:

Citation rate = 123 ÷ 560 ≈ 22%

That 22% applies only to the chosen questions, engine, location, and collection method, never to every ChatGPT conversation about product analytics.

How sure is 22%? Assuming independent answers, 123 citations in 560 answers give a 95% margin of about 3.4 points. The chart shows that margin shrinking as answers are added.

A change between two windows has its own, wider margin. For two independent windows of 560 answers at about 22%, the 95% margin around their difference is about 4.9 points, about 1.4 times either rate's margin.

A frozen question set breaks the independence assumption, since questions differ and repeated runs depend on each other. A real panel therefore needs an interval built for that design, such as a bootstrap that resamples whole questions with their runs.

Some tools divide by all citations instead, which gives citation share, so every report needs its formula next to the number. In Atomic, the AI Visibility view compares citation share by engine, and Recent chats opens each saved answer with its citations for a check before reporting.

What to do: Report citation rate per engine and cluster with its answer count, formula, and uncertainty method. Judge a trend by the change interval, and check a factual error or a sudden loss whenever it appears.

2. Mention rate and sentiment: how AI answers describe the brand

Mention rate is the share of answers that name the brand. The useful version also records each answer's brand sentiment: recommended, listed, or warned against.

In the Semrush study, ChatGPT linked an appearing brand 87% of the time but named it in only about a fifth of appearances, and Gemini did the reverse. Each engine needs its own row in the report.

Atomic's Prompts page labels each prompt's ranked inclusion High, Medium, or Low. The Missing filter in Recent chats shows answers without a ranked brand position, even when the text names the brand in passing.

Atomic's sentiment view weights each positive or negative observation by its assessed impact, from 1 for a minor signal to 3 for major direct evidence. Its positive share is the positive weight's share of all assessed weight, and neutral answers stay out of it.

Atomic's documentation sets a threshold for acting: "One negative mention is noise. Repeated negative themes across prompts are the real risk." A recurring warning sets the order of investigation: open the evidence behind the theme, correct a verified product error, and check the cited pages for outdated facts.

What to do: Count mentions and citations in separate columns, record the framing beside each mention, and check automated sentiment against answers read by hand.

3. AI share of voice: the brand's share against named rivals

AI share of voice is the brand's share of the mentions, or citations, that the brand and a fixed set of rivals receive in the same answers. A citation rate alone cannot show whether the brand is gaining share against rivals, because their rates can rise in the same month.

In the worked example, Vendor A's share of mentions against four fixed rivals rises from 24% to 27% on the same 20 comparison questions. Judging that rise takes the answer records behind it and an interval that keeps the repeated questions together.

Atomic's Competitors page puts the brand and its named rivals side by side, engine by engine, with visibility, average position, and citations for the same prompts and dates. Formulas differ between tools, and some weight each answer by a modelled search volume. Two tools' shares compare only when their setups match, from the questions and rival set to the counting rules. The guide to AI share of voice covers how to choose the rivals.

What to do: Fix three to five named rivals per cluster, keep mention-based and citation-based share of voice apart, and read the share once a month.

4. Core sources: the pages AI engines keep citing

Most cited websites rotate from week to week, while a few keep coming back. SISTRIX calls that stable group the core and the rest the carousel. In its data, 86.5% of Google AI Mode prompts had a stable core of one to five domains.

Atomic's own tracking looks the same. Across all its tracked engines, after merging duplicate URLs, 60% of the 8,698 links cited for its questions over 30 days appeared only once.

On a single question, the core gets sharper. In 12 Google AI Overviews for one of Atomic's buying questions in October 2026, a comparison list on a tool vendor's blog appeared in all 12. That list is the first page to check for an outreach request.

Allsopp's daily checks show the same split. In ChatGPT, the websites cited on at least 80% of days changed 14% from one day to the next, against 74% for the rest.

This guide counts a page as a core source when it appears in at least half of the runs for a question or cluster, its own convention. Atomic's Citations page ranks cited pages and domains by share and labels each by type, so the core sources arrive sorted by owner.

What to do: Review the core sources per cluster each month, sorted by owner, and start with the third-party pages that skip the brand and fit the product.

5. AI referral traffic: the visits that arrive

AI referral traffic counts the sessions and key events, or conversions, that GA4 attributes to answer engines. It cannot say which tracked answer sent a visit, and it misses every answer that sent no click.

It also misses visits that arrive without a tag or referrer, as the ChatGPT tagging break of October 2026 showed, so referral reports undercount AI visits. In Atomic, the AI search Overview shows AI clicks, which are GA4 sessions from AI platforms, with their conversions by engine and page. Its Pages view lists the URLs with AI clicks beside the Search Console pages that have none. Its documentation adds: "Always compare sessions and conversions together. Volume without outcomes is not a growth signal."

What to do: Read AI referral sessions and key events by landing page each week, read Search Console by page group, and treat referral counts as a floor.

6. Citation decay: when a page loses its place

Citation decay is the fall in a page's citation rate from one window to the next, on the same frozen questions. Its speed depends on what is measured. Following each page's share from its peak, Profound found that the median page lost half within 11 days, while Scrunch's weekly groups of cited sources took four to five weeks to halve.

Atomic's Citations page charts citation share by domain over time and ranks every cited page in its By page tab. A page's decline is diagnosed at URL level, because a domain can stay cited while its cited page changes.

In the worked example, Vendor A's pricing guide falls from 22% to 14% to 9% over three 28-day windows of 140 answers each. Assuming independent answers, the first fall of 8 points sits inside a margin of about 9. The 13-point fall from the first window to the third clears its margin of about 8. The answers now cite a newer rival comparison, which Vendor A reads before refreshing the guide.

What to do: Compare 28-day windows on a frozen question set and flag pages whose fall clears the change interval. Read a sudden loss on a high-value question anyway, and check what replaced the page before a refresh.

How to track AI citations step by step

Tracking starts with a stable set of real buying questions. The run budget and the reporting window follow from the size of change the team needs to detect.

A quick check by hand shows the problem first: ask one buying question in five fresh, logged-out chats, save each answer with its cited URLs, and count the pages that appear in all five. ChatGPT tracking tools automate the same check at the volume a rate needs.

Atomic began inside Omnius, and Omnius's own video walks through how these tools collect answers and why their visibility scores are estimates:

Seven steps turn that check into a measurement:

1. Start with the data the website already has. Use the AI Assistant channel or the custom channel group below, read Search Console's generative AI report, Bing Webmaster Tools' AI Performance report, and Microsoft Clarity's Citations dashboard, and count visits by ChatGPT-User, Claude-User, and Perplexity-User in server logs. In Atomic, connecting GA4 and Search Console in Data sources fills the AI search Overview with GA4's AI sessions, and Crawler logs, on Team plans and above, list AI bot requests and failed responses. Each measures something different: GA4 counts visits, Search Console counts impressions, Bing and Clarity count citations, and server logs count page fetches. In Search Console, a query whose click-through rate falls while its position holds is a lead to inspect, so check whether an AI Overview now appears for it before blaming the Overview.
2. Write the question set. Take questions from sales calls, support tickets, Search Console question queries, Reddit threads, and prompt research tools, in customer wording with real constraints. Check that the answers address the intended meaning before freezing a question: in Atomic's five answers, half of the 16 citations went to pages about tracking what ChatGPT itself costs to run, a second reading of the same words. Group the questions by intent, keep branded ones in their own cluster, and freeze the set with a version number. In Atomic's Manage prompts, each line becomes a prompt with a funnel stage and a group, and the groups sidebar shows the slots in use.
3. Fix the engines and conditions. Choose engines by where customers ask, starting with the assistants that already send AI referrals in GA4, then fix country, language, logged-out state, mode, and collection method for every run. Atomic's AI search engines control sets which of the plan's engines run each prompt.
4. Set the runs from the margin needed. For the level of a rate across customer questions, broader question coverage reduces the uncertainty that comes from picking questions, which repeat runs of the same questions cannot remove. For a change on a frozen set, keep the questions and their weights fixed, and precision then depends on the runs per window and how much they vary. Give the set a new version number whenever questions are added.
5. Store every answer. Keep the date, engine, conditions, question version, run status (completed, failed, or pending), every source URL in order as shown and in normalised form, whether each link is a citation or only a related link, every brand named with its framing, and the full text. Atomic stores completed answers with their citations and ranked brand results, and its mention percentage leaves failed, pending, and unclassified answers out. Report pages and domains side by side, because in SISTRIX's data the cited pages changed more from week to week than their domains.
6. Sort cited pages by owner each month. Label each as own, rival, or third party, and list the core sources; Atomic's Mentions page splits the cited third-party pages into those that mention the brand and those that do not.
7. Keep a change log and a control set. Log engine releases and page edits with their dates, and judge each change against questions or pages that were left alone.

Google's own sample regex matches any source ending in ai or containing gpt, which pulls in unrelated websites, so the rule above lists exact domains.

How one tracked question becomes a 28-day report

In the worked example, Vendor A tracks 40 buying questions, 20 on pricing and 20 on comparisons, once a day on ChatGPT, Google AI Overviews, Google AI Mode, and Perplexity. A 28-day window holds 1,120 runs per engine.

Pricing question P07 asks: What does product analytics software cost for a 50-person B2B SaaS company that needs a data warehouse sync? Its first ChatGPT answer, on 14 September 2026, cites five pages and carries the warehouse-sync warning quoted earlier.

Over 28 runs, 9 answers cite a Vendor A page, a 32% rate with a margin of about ±17 points under the same assumption. That margin makes one question a poor guide to change, so the report compares the cluster rates: 22% for the 20 pricing questions and 21% for all 40.

The 28-day report fits on one page, with ChatGPT sessions from 8 October marked for the tagging break that began that day:

An Overview appeared on 812 of the 1,120 runs, so the 26% rate covers only the answers that existed. Across all runs, a Vendor A page was cited about 19% of the time. The table counts completed runs, and its margins assume independent answers. Behind the ChatGPT row sit 235 citing answers, 157 that name Vendor A, and 654 mentions of the five tracked brands.

What are the most common AI citation tracking mistakes?

Even a careful report misleads when read the wrong way, in five common ways.

Reading a daily swing as a trend

A six-point jump in visibility from one day to the next looks like news. Yet 50 prompts a day give only 50 answers. For two independent days of 50 answers at about 20%, the margin around the difference is roughly 16 points, so a six-point jump sits inside that interval. The panel's own change calculation decides at every scale.

Reading answer order as a ranking

It is tempting to read the order of an answer's citations as a rank. In Allsopp's data, the domain ChatGPT cited most for the typical prompt appeared in 82% of daily answers yet came first on only 43% of days.

A position in a report needs its definition: citation order, a place in a ranked list, or a dashboard's rank by share. Absence needs its own column too, since Atomic's average position, for one, counts a missing answer as position 10.

Mixing branded and unbranded questions

The mix of question types moves every rate. In the Semrush study, comparative prompts named the brand more than twice as often as informational ones.

Suppose Mixpanel tracks best product analytics tools for B2B SaaS beside Is Mixpanel worth it for a startup? The second names Mixpanel by design and inflates the mention rate. Branded questions belong in their own cluster.

Crediting traffic changes to citations

Referral data can break for reasons unrelated to citations. From 8 October 2026, the day after GPT-6 began rolling out in ChatGPT, many ChatGPT links carried no tracking tag and no referrer. Analytics filed those visits as Direct.

Cedric Huesler of Adobe compared 9 to 13 October with the week before on about 500 pages where ChatGPT drove 5% or more of visits. ChatGPT visits fell by about 31,000 a day and Direct rose by 26,000, while search moved like everywhere else. His "first guess was that GPT-6 shows fewer links."

Huesler reported the tag back on 15 October. By the next morning, about 99% of Profound's collected answers with cited links carried at least one tagged source link. A referral change on its own cannot show whether a citation won or lost a visit.

Claiming a cause without a control

Without a control, a before-and-after comparison works like timing a runner on a moving walkway: the time improves whether or not the runner did.

In Ahrefs' test of 1,885 pages that added schema, AI Mode citations rose 43%. About 4,000 matched control pages that added no schema rose almost as much, so the schema effect shrank to 2.4%, which Ahrefs could not tell from zero.

Engines also move the baseline in public. On 15 October 2026, Google said it had made "continued and substantial increases" to inline links in AI Overviews and AI Mode, so a citation count can rise with no change on the website.

How to use citation data to improve AI visibility

A citation report earns its place when it changes the team's next job, starting from the core sources.

Close the third-party source gaps

In the worked example, Vendor A's pricing question shows the sorting. Over 28 runs, three pages appear in at least half of the answers, and Vendor A's own page trails them:

• A review website's category page appears in 22 of 28 runs and skips Vendor A, so the first job, once the category fits, is a listing request.
• A rival's pricing page appears in 18 of 28 and answers the same question as Vendor A's own pricing page.
• A Reddit thread appears in 15 of 28 and repeats the old warehouse-sync limit, so a disclosed reply with the current plan corrects it.
• Vendor A's pricing page appears in 9 of 28 and states the warehouse sync only inside an image.

A comparison list that engines keep citing, and that leaves the product out, is a specific outreach target when the product fits its criteria. Atomic's Mentions report checks cited third-party pages for brand coverage and missing opportunities. Since 13 October 2026, a beta tab in Bing Webmaster Tools also lists opportunity queries, where other websites are cited and the website is not.

Learn from the rivals' cited pages

The rival's pricing page, cited in 18 of 28 answers, lists its plans and integrations in the page text, with a date, while Vendor A's page keeps the same fact inside an image.

Fix the facts on the brand's own pages

Google recommends making important content available in textual form, so Vendor A moves the warehouse-sync detail into the page text, with a date. The same pass checks each engine's own conditions:

• Google AI Overviews and AI Mode: Google's AI features guide asks for a page that is indexed and eligible to show a snippet in Search.
• ChatGPT search: OpenAI asks websites to allow OAI-SearchBot and requests from its published IP ranges, and says ChatGPT search can take about 24 hours to adjust after a robots.txt change.
• Claude and Perplexity: Claude-SearchBot indexes pages for Claude's search results and PerplexityBot surfaces websites in Perplexity's answers, so each needs access in robots.txt and in the content delivery network's bot rules.

Atomic's technical SEO audit checks that access for Google, AI crawlers, and LLM agents, along with rendering, indexability, and schema.

Bing's grounding queries supply a second list of fixes. Relevant sampled searches that a page leaves unanswered are leads, and the confirmed gaps belong in the next content brief.

Each fix is a test. Vendor A compares the changed pages with similar pages it left alone and checks the timing, other edits, and engine updates before crediting the fix.

Keep a working rhythm

A weekly read catches broken tags and sudden drops. A monthly report covers the rates, share of voice, and core sources against the control set, and a quarterly review turns the citation gaps into the next content and outreach plan.

Atomic can email a weekly digest and a monthly performance report, and its agents can run the recurring parts on a schedule. The guide to AI search monitoring tools covers that routine.

Best tools and methods for AI citation tracking

A check by hand suits a small diagnostic panel and can catch a wrong fact in one answer. Software volume follows from questions, engines, and runs. The worked example schedules 4,480 runs in 28 days, and answers are counted apart, since a Google run can return no Overview. The list of AI search tracking tools compares the software side by side.

An SEO suite's add-on suits a team extending its SEO reports to AI answers, and a dedicated tracker suits AI visibility run as its own program.

Four questions for any citation tracking tool. Does it query the consumer app or the developer API? Which country and account state does it use? How many runs per question does it make in a window? Does it store the full answer text and every cited URL?

How Atomic tracks AI citations

Atomic AGI treats AI citation tracking as sampling, because "AI responses naturally vary between runs." The same questions run on a schedule on the project's selected engines, set by the plan, through the AI products' interfaces as well as their APIs, and every answer is stored beside GA4 and Search Console. Atomic is free to start, and the pricing page lists the plans.

Atomic's own video shows the same tracking against named competitors:

See how Atomic's AI search tracking records citations and mentions for a fixed question set, beside GA4's AI sessions.

Frequently asked questions

Can ChatGPT find citations?

Yes, ChatGPT can find and show citations when it searches the web, and its Sources button lists them when available. OpenAI also warns that search results and citations "can be incomplete, outdated, or incorrect."

Can I use AI to find citations?

Partly. AI search can show source links one answer at a time, though not every answer includes them. Finding where a brand is cited takes repeated runs, which tools such as Atomic automate.

Is there an AI for citations?

Yes, there are AI tools for citations. AI citation tracking tools run fixed questions across AI engines on a schedule and store the cited pages and the brands named, and Atomic is one.

What is citation rate (citation frequency) and how is it calculated?

Citation rate, also called citation frequency, is the share of tracked answers that cite a page or domain. Citation share divides by all citations instead, and some tools use the name for other ratios.

Do citations from AI tools drive referral traffic?

Citations from AI tools drive some referral traffic, but an answer read without a click leaves no referral record. In March 2025 browsing data, Pew Research Center found that Google users clicked a link inside an AI summary in 1% of visits to results pages that showed one. Google's default channel group guide explains how GA4 classifies the AI visits that do arrive.

Can Google Search Console show which pages get cited in AI Overviews and AI Mode?

Google Search Console shows which pages get cited in AI Overviews and AI Mode only in part. Its generative AI report shows their impressions without clicks or a query filter, and in the main Performance report every link in an AI Overview shares one position.

Can AI citation tracking show what customers ask?

No, a citation tracker describes the answers it collected, never the questions customers asked elsewhere. Bing Webmaster Tools shows a sample of grounding queries, the searches Microsoft's AI ran, which are not the customers' own prompts. Prompt datasets in some tools offer sampled or modelled demand.

Methodology

This guide combines platform documentation, outside studies, and Atomic's own tracked questions:

• How the conclusions were reached. The guide rests on documentation from Google, OpenAI, Anthropic, Perplexity, and Microsoft, read on 17 October 2026 and rechecked on 20 October 2026, with OpenAI's help pages partly read from archived copies where automated reads failed, and on the outside studies in the table below. Atomic's tracker, which collects answers through the AI products' interfaces as well as their APIs, gave 12 stored Google AI Overviews for one of its buying questions about AI tracking tools from 2 to 16 October 2026, and five stored ChatGPT answers for another from 18 to 28 September 2026.
• Counting rules. Cited URLs were counted as stored, after lowercasing the host and removing the scheme, query string, fragment, and trailing slash, and each answer counts a page once. Page types come from the cited URLs, read on 17 October 2026.
• Limits. The tracked questions are Atomic's own buying questions, chosen instead of sampled, so the checks describe those questions only. The ChatGPT answers sat two to four days apart. Nozzle, a rank-tracking company, reported that from 20 September 2026 Google served rank trackers results pages with their usual websites missing, and in its update of 19 October those pages showed an AI Overview 45.9% of the time against 53.6% in its baseline. The 12 stored AI Overviews fall in that window, and the guide draws no conclusion from runs that showed no Overview. The checks count shown links only and make no claim about why a page was chosen.
• Worked example. Vendor A, its questions, answers, pages, and report figures form a worked example constructed for this guide, counted one mention per brand per answer. Its margins use the normal approximation for a proportion, 1.96 × √(p(1 − p)/n), and for a change between two windows the same formula for the difference of two proportions, with answers treated as independent. A real panel has question-level differences and repeated runs, so its interval needs a method that keeps each question's runs together, such as a question-level bootstrap. NIST also presents the Wilson interval, whose width is close to the normal one at a few hundred independent answers.
• Images. The screenshots of Google's, Microsoft's, and OpenAI's pages were captured on 17 October 2026. The ChatGPT answer comes from Omnius's own signed-in account in October 2026, so personal context may have shaped it; it shows what a mention and a citation look like, and the client result it quotes and the other recommendations are blurred. The Atomic views are Atomic's own product images, shown on an example project. The videos are Atomic's own and Omnius's own, the latter published on 23 June 2025.

The outside studies, with each publisher's interest:

Cedric Huesler works at Adobe, Nozzle is a rank-tracking company, and Search Engine Roundtable and Search Engine Journal are trade publications, cited for reported dates and statements.

Sources

The documentation, studies, and reports behind this guide:

• Google documentation: AI features guide, Search Console metrics, generative AI report, AI Mode help, why results differ, Gemini app sources, default channel group, custom channel groups, Analytics release notes, Gemini API grounding
• Google announcements: Search Console AI reports, controls for website owners, inline links, May 2026, inline links, October 2026
• OpenAI, Anthropic, and Perplexity documentation: ChatGPT search, OpenAI publisher FAQ (archived), OpenAI crawlers, OpenAI web search API, Anthropic crawlers, web search in Claude, Claude web search tool, Perplexity crawlers, how Perplexity works, Perplexity Pro Search, Perplexity Agent API, Memory in ChatGPT
• Microsoft: AI Performance launch, Citation Share and Intents, Copilot search, Clarity Citations, Topics Portfolio
• Studies and vendor reports: SISTRIX, Detailed, Ahrefs on the top 10, Ahrefs schema test, Ahrefs Brand Radar method, Ahrefs AI demand, Ahrefs AI metrics, Ahrefs AI volume, Semrush ghost citations, Semrush AI metrics, Semrush AI data, Profound half-life, Profound on missing UTM tags, Profound glossary, Scrunch half-life, Nozzle on poisoned results pages, Pew Research Center, NIST proportion intervals, NIST difference of proportions
• Atomic and Omnius: prompts, metrics, citations, sentiment, competitors, attribution, data sources, technical SEO, AI agents, about Atomic, Omnius video on tracking tools
• Trade press and posts: Search Engine Roundtable on tags, on ChatGPT sources, on Search Console, Search Engine Journal on GPT-6, on Gemini tags, Digiday, Cedric Huesler on LinkedIn, Search Engine Journal on rank trackers