AI Search Competitor Analysis: How to Find Out Why AI Recommends Your Competitors Instead of You

You open ChatGPT, ask a buying question, and a competitor’s name appears three times while yours appears once - or not at all. The reflex is to assume a content gap: they wrote more, they published fresher, their domain has more authority.

You may have spotted the symptom and missed the cause. Follow that assumption without checking it, and you can spend a quarter fixing the wrong problem.

The question you need to answer isn’t “how do I rank higher.” It’s “what is the machine actually using to justify this verdict?” An AI answer isn’t a ranking. It’s a generated summary assembled from a set of retrieved pages. Every brand in that answer earned its place because some specific page said something. Find that page, and you have found the reason. Find out that the same class of page is carrying three competitors and not you, and you have found your actual work.

This guide is a method for doing exactly that: run a controlled prompt set, capture the claims, trace each claim to its source, classify what kind of source it’s, and convert the pattern into a specific fix.

There’s one hard truth that shapes everything below, so it comes first.

The engines will not show you the competitor - and that is deliberate

In June 2026 Microsoft added Citation Share to the AI Performance report in Bing Webmaster Tools. It’s the closest thing any engine has shipped to a competitive AI-visibility metric. Here is what it is, verbatim from the announcement:

“Citation Share shows how much of the citation space your site receives for a specific grounding query. It is calculated as the percentage of citations attributed to your site out of all citations shown across all sites for that same grounding query.”

And here is the part that matters:

“Importantly, Citation Share is designed as an observational metric – not a ranking system or a competitive scoreboard. It does not expose competitor domains, represent traffic share, or assign quality scores to content.”

Microsoft built a citation-share metric, and in the same paragraph specified that it wouldn’t tell you who the other side of that share belongs to.

Four months later Microsoft shipped the closest thing to a competitive gap tool I have found, and it made the same choice in a subtler way. Topics Portfolio (Beta), announced October 13, 2026, groups the grounding queries where your site appears into thematic clusters. It has a column called Sites Cited, defined as “the approximate number of sites appearing for queries in this topic, which indicates how crowded the topic is.” And it has Top Opportunity Queries - “the strongest queries across your whole portfolio where other sites appear and yours does not.”

Read that again. Microsoft will tell you the queries where other sites appear and you don’t. It will tell you how many other sites appear. It won’t tell you who they are. The post is explicit that “Everything in this view is personalized to your site. The topics and the opportunity queries are surfaced from your own presence and activity in AI Search, not from a generic ecosystem list.”

Microsoft also frames Topics Portfolio for exactly your use case - “it should be most useful to publishers doing content strategy, editorial planning and competitive positioning” - while still withholding the competitor names.

Google’s side of the ledger is the same shape. The Generative AI performance report in Search Console, described in Google’s AI optimization guide, measures “how your content is performing in generative AI features on Google Search and Discover.” It’s a report about your content.

Any cross-brand “AI share of voice” number you see marketed anywhere has to be computed by somebody - either that vendor, or a commercial tracker. Your options are to pay for one of those, or to compute it yourself. This guide is about computing it yourself, in a way that survives scrutiny.

A second limitation: citation share is not a scoreboard, and “Compare” is not what it sounds like

Bing also shipped a feature called Compare alongside Citation Share. It isn’t a competitor-comparison feature. It “allows publishers to overlay a previous time period directly onto the current reporting view” - a 30-day window against the prior 30. Anyone building an AI competitive dashboard who reads “Compare” in a release note and assumes it means “compare me to my rivals” has misread it.

Also worth knowing before you over-read any single metric: the February 2026 AI Performance announcement is careful that Average Cited Pages “reflects overall citation patterns and does not indicate ranking, authority, or the role of any page within an individual answer,” and that Total Citations shows frequency “without indicating placement or presentation within a specific answer.”

Frequency isn’t influence. Track it, but don’t let it stand in for being recommended.

Why classic rank tracking cannot answer this question

The instinctive move is to check whether the competitor outranks you in classic search for the same terms, and if they do, conclude that’s the cause. A peer-reviewed measurement study argues against that inference directly.

Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact, accepted to ACM Internet Measurement Conference (IMC 2026) and issued as arXiv:2605.14021 (v2, October 1, 2026), ran 55,393 trending queries across 19 topical categories over a 40-day window from March 13 to April 21, 2026. One of its four headline findings:

“AIO-cited domains are more credible than co-displayed first-page results, yet nearly 30% do not appear in those results at all, indicating a source selection mechanism distinct from Google’s ranking algorithm.”

Nearly a third of the domains Google cites in AI Overviews aren’t on page one of the classic results beside them. Start with a rank tracker and you’ll miss nearly a third of the sources the engine used. That leaves you under-counting competitors who do well in AI retrieval and poorly in classic SERPs. You can end up spending more on SEO fixes that don’t address the gap.

The same paper gives you a second reason to distrust shortcuts. Decomposing responses into 98,020 atomic claims, it found that 11.0% are unsupported by the cited pages, with omission the dominant failure mode. In other words, in roughly one claim in nine, the citation displayed next to a statement doesn’t actually contain that statement.

That’s why I want you to open the source page. In roughly one case in nine, the page cited beside a claim doesn’t contain that claim. Record the URL without reading it, and your spreadsheet carries the error forward.

Step one: build a prompt set, because the prompt you type is not the query that runs

A controlled competitor analysis starts before you look at a single answer, and the first thing to get right is that engines don’t answer the text you typed.

OpenAI’s web search documentation describes three search modes, and in the agentic mode “the model actively manages the search process. It can perform web searches as part of its chain of thought, analyze results, and decide whether to keep searching.” Responses include a web_search_call item whose action may be search, open_page, or find_in_page. Your single sentence can become several searches, a page fetch, and a within-page search.

The ChatGPT search help article makes the rewriting explicit:

“ChatGPT search typically rewrites your query into one or more targeted queries that it sends those providers.”

It gives the pattern: a user asking about the latest on a drug class might first be sent a broad query, then, “after reviewing the initial results,” a more specific one naming a particular compound.

Google does the same thing and calls it query fan-out. From the AI optimization guide (last updated 2026-07-10 UTC): a set of concurrent, related queries generated by the model, where a user query of “how to fix a lawn that’s full of weeds” might fan out to “best herbicides for lawns,” “remove weeds without chemicals,” and “how to prevent weeds in lawn.”

Bing surfaces the result of this as grounding queries - “the key phrases the AI used when retrieving content that was referenced in AI-generated answers” - and as of June 2026 classifies them into Intents (Informational, Commercial, Navigational, Learn and Solve, Research, Creation, Local, and more) and Topics (thematic clusters, e.g. “Solar Energy”).

What this means for your method. If you run one prompt once and record what came back, you have measured noise. Build a set instead:

• Fix the engine, fix the geography, fix the date. Record all three on every row. Location matters because ChatGPT “may use an approximate location based on your IP address to provide relevant local results,” and OpenAI’s API accepts an explicit user_location. A competitor analysis run from two countries is two different analyses.
• Separate prompt intent from engine internals. Group your prompts by what a buyer would actually ask, then let the engine fan out. You want to measure the outcome, not reverse-engineer the rewrite.
• Include the branded and the unbranded. A prompt naming no vendor tells you who the engine volunteers. A prompt naming your competitor tells you whether you survive head-to-head. These produce different remediation.
• Repeat. Microsoft notes citation patterns “can shift due to changes in user behavior, evolving models, freshness signals, partner refresh cycles.” Any single run is a sample. Record n and date.

Step two: capture what was read, not just what was shown

This is where most competitor analyses quietly break. You record the inline citations the interface displayed, and you assume that’s the evidence base.

It isn’t. OpenAI’s documentation is explicit:

“To view all URLs retrieved during a web search, use the sources field. Unlike inline citations, which show only the most relevant references, sources returns the complete list of URLs the model consulted when forming its response. The number of sources is often greater than the number of citations.”

The citations you can see are only part of what the engine retrieved. A competitor’s page might have been consulted without being shown. If we count visible citations alone, we miss that contribution.

Two ways to widen the capture:

In ChatGPT itself, the help article notes you can “Select Sources, when available, to view cited sources and other relevant links.” Use it every time, and record it as a separate column from the inline citations.

In the API, request include: ["web_search_call.action.sources"] to get the full consulted-URL list programmatically. OpenAI also documents domain filtering - “up to 100 allowed_domains or up to 100 blocked_domains” on the Responses API web_search tool - which turns the prompt set into a genuine experiment. Blocking a competitor’s domain and re-running tells you what the answer looks like without them.

I’d pay attention to that last test. Run the prompt unrestricted, then with the competitor blocked, then with your own domain allowed. Compare what changes in that answer. This is the closest we can get here to testing the competitor’s contribution, rather than counting their citations.

The same documentation also sets a limit worth internalizing: search_context_size “does not set an exact token count or guarantee a specific number of sources or citations.” Answers vary. Treat every cell in your grid as an observation with a date attached.

If your competitive work is on Perplexity, the crawler rules are the prerequisite. Perplexity’s crawler documentation states that PerplexityBot “is designed to surface and link websites in search results on Perplexity” and is “not used to crawl content for AI foundation models,” with Perplexity-User handling user-initiated requests. OpenAI draws the same line: OAI-SearchBot “is for search” and used to “surface websites in search results in ChatGPT’s search features,” while GPTBot is for training, and “[s]ites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers.” Anthropic splits three ways - ClaudeBot for training, Claude-SearchBot for search indexing, Claude-User for user-initiated retrieval.

Before you conclude a competitor “beats you in AI search,” confirm they haven’t simply blocked the crawler you’re measuring on. That’s a teardown finding in its own right.

The Recommendation Source Anatomy: classifying what is carrying the winner

Here is the analytical core of this method.

Google’s AI optimization guide, Microsoft’s AI Performance work, and OpenAI’s web search documentation don’t give weights by source type. We can’t fill that gap with a made-up scoring formula.

What you can measure is the composition of source slots - which kinds of pages occupy the citations in the answers where competitors win. That’s a fact you can record.

Six source classes cover almost everything that shows up:

Class	What it is	Typical owners	Why an engine might reach for it
A - First-party	Your site, docs, pricing, changelog	You	Only source with the most authoritative version of a factual claim about itself
B - Reference	Wikipedia, dictionaries, standards bodies	Third party, neutral	Stable, structured, cheap to parse, low drift
C - Community	Reddit, Quora, forums, Stack Overflow	Third party, adversarial	Real-world experience; huge volume; captures complaints and edge cases
D - Review & trust	Consumer Reports, J.D. Power, BBB, Trustpilot, G2-style review sites	Third party, gatekeeping	Compressed multi-source judgment on quality
E - Editorial	Roundups, “best X in 2026” listicles, comparison pages, trade press	Third party or competitor-adjacent	Directly matches commercial-intent prompts
F - Transactional & integration	Retailers, marketplaces, app stores, integration/MCP registries, agentic commerce feeds	Third party, structured	Normalized structured data a system can consume without parsing prose

Google’s own documentation names several of these classes without using my labels. Its AI optimization guide notes that its generative AI features “can show what’s being said about products and services across the web, including in blogs, videos, and forum discussions” - classes E and C. Google’s people-first content guidance (last updated 2026-10-05 UTC) counts “Forum discussions, customer reviews, comments, or user-uploaded media” within legitimate main content when they serve the page’s purpose.

And the composition of AI answers is genuinely different per engine. Ahrefs published two parallel studies in September 2026 using one methodology and dataset family across a broad set of US queries, via its Brand Radar product. Treat these as vendor research, not platform truth - but the cross-engine contrast is the point:

Google AI Overviews, top mention share: youtube.com 22.9%, reddit.com 18.5%, facebook.com 10.1%, google.com 8.8%, instagram.com 5.6%, quora.com 4.7%, en.wikipedia.org 4.0%.

ChatGPT, top mention share: reddit.com 16.8%, en.wikipedia.org 7.0%, consumerreports.org 3.7%, forbes.com 3.1%, engineerfix.com 2.8%, legalclarity.org 2.8%, healthline.com 2.5%, youtube.com 2.5%.

Two things jump out.

First, YouTube’s share collapses by roughly a factor of nine between the two - 22.9% of AI Overviews citations versus 2.5% of ChatGPT’s. A competitor whose dominance is built on video will look like a rival to beat in one engine and an irrelevance in the other. If your entire competitive read comes from Google, you’ll misjudge the field.

Second, ChatGPT’s list carries class D and class B domains that barely register in AI Overviews - Consumer Reports at 3.7%, J.D. Power, BBB, Trustpilot, TripAdvisor, plus dictionaries and government and medical sources like FDA, CDC, MedlinePlus and PubMed. AI Overviews’ top ten is dominated by social and video platforms. If your buyers ask trust-shaped questions, the competitive set on ChatGPT is a completely different set of companies than the competitive set on Google.

A caveat that matters for reading any such dataset: Ahrefs defines mention share as “a domain’s citations as a percentage of the summed citations of the top sources” - relative to the top 50, not share of all citations. The data is US-only. And as noted, these pages are “generated and automatically kept up to date every month by Agent A,” so figures are a monthly snapshot.

Computing the Anatomy Index

For each prompt in your set, classify every source slot in the answer by class A–F. Then, per competitor:

1. Count source slots occupied by each class in answers where they appear.
2. Repeat for answers where you appear.
3. Compute the gap per class.

The output is a one-line diagnosis. Something like: “Competitor A holds 6 of 9 community-class slots in commercial-intent prompts; we hold zero and have no presence in that class.”

That sentence tells you what to build. A visibility score doesn’t. It also helps you pick a fight you can win. If a competitor wins on class C and you have no way into class C, focus on class A and D, where you can write the source yourself. Chasing their apparent weakness could be expensive and get you nowhere.

Prioritization rule: weight remediation toward classes where (a) the gap is large, (b) you have genuine and defensible material to contribute, and (c) the class is one the engine demonstrably uses. Don’t weight by class prestige. On a query the engine resolves from lived experience, a perfectly executed page on your own blog is unlikely to out-rank a consolidated community thread.

The competitor teardown worksheet

One row per prompt × engine × date × geography. This is the working artefact; the write-up is downstream of it.

Identification

• Prompt ID, exact prompt text, prompt intent class (commercial / informational / comparison / navigational)
• Engine and exact surface (AI Overviews, AI Mode, ChatGPT search, Copilot, Perplexity)
• Date, time, and geographic location of the run
• Whether the engine ran a web search at all (ChatGPT may answer without one)

The recommendation

• Was each brand mentioned? Position in the list if ordered.
• The exact claim attached to each brand, quoted verbatim - not paraphrased. Paraphrase destroys the only evidence that matters.
• Sentiment and specificity: is the claim a differentiator or filler?

The evidence behind it

• Source URL, displayed as citation or discovered via sources / “Sources”
• Page owner: you / them / third party
• Source class: A–F
• Did the page actually support the claim? (verified / partially / not supported)

That last field is the one teams skip, and the IMC 2026 finding - that 11.0% of atomic claims in AI Overviews were unsupported by their cited pages, with omission dominant - is the reason not to. OpenAI’s own help article makes the same warning from the other direction: “Search results and citations can be incomplete, outdated, or incorrect. Open a cited source to check that it supports the answer.”

The diagnosis

• Where do you and the competitor differ in class composition?
• Is the competitor’s winning page indexed and eligible, or is it winning despite a defect you can exploit?
• Are they blocked, noindexed, or canonicalized elsewhere? Is their content duplicated across their own site?
• Is the claim about product facts (class A territory, contestable) or about reputation (class C/D territory, expensive)?

The action

• Remediation mapped to a source class, an owner, and a date
• Explicitly: what you’re not going to do, and why

The last line matters. A teardown that generates only action items will produce a content treadmill aimed at a moving distribution. Recording what you’re deliberately not chasing is what makes the rest credible.

Turning the diagnosis into fixes that engines actually document

Once you know the winning source class, the remediation is mostly determined. Here is what has first-party support, with dates.

If the gap is class A - first-party. Google’s AI optimization guide is the current canonical statement (last updated 2026-07-10 UTC). Its preferred framing is “non-commodity content”: it offers “Why We Waived the Inspection & Saved Money” as a model of content with “unique expert or experienced takes that go beyond common knowledge.”

There’s an administrative prerequisite that’s easy to miss. The same guide states: “In addition to the technical requirements for Search, a site must be included in Search generative AI features in Search Console to be eligible for display in generative AI features on Google Search.” Google’s earlier AI features page, last updated 2025-12-10, says instead that “There are no additional technical requirements.” These aren’t reconcilable as simultaneous truths. The most defensible reading is that the Search Console inclusion step is newer and the 2025 page predates it; if you’re eligible today, the 2026 gate is the one to satisfy. Check it before you audit anything else.

If the gap is class D - review and trust. Google’s product structured data guidance (last updated 2025-12-10) distinguishes Product snippets - “For product pages where people can’t directly purchase the product… this markup has more options for specifying review information, like pros and cons” - from Merchant listings for transactional pages. Review markup on non-purchase pages is a specific, documented lever for how your product page can surface in Search. Be precise about what it buys: Google’s documentation describes these enhancements for search results, and separately warns that structured data is “not required for generative AI search, and there’s no special schema.org markup you need to add.” It’s a defensible hygiene fix, not a lever with a documented AI payoff.

If the gap is class A plus a distribution advantage - preferred sources. Google’s preferred sources guide, last updated 2026-09-18, documents a lever most competitors haven’t noticed: “In AI Mode and AI Overviews, your content can be highlighted with a ‘preferred’ badge for users who have selected your site as a preferred source.” Only domain- and subdomain-level sites are eligible - https://www.example.com/blog isn’t. Implementation is a two-line script tag plus a div:

<script async src="https://news.google.com/swg/js/v1/publisher.js"></script>
<div google-add-preferred-source-btn></div>

Or, with no JavaScript, a deeplink to https://www.google.com/preferences/source?q=example.com.

This is the single most directly teardownable item on the list. Run curl -s https://competitor.com | grep -c publisher.js. If the competitor has the button and you don’t, that’s a concrete, documented, free gap.

If freshness is the differentiator. IndexNow is the documented mechanism both Microsoft and Google point to. Microsoft’s AI Performance announcement says “IndexNow helps ensure that AI systems reference the most current version of your page when generating answers.” The IndexNow documentation describes pushing up to 10,000 URLs per POST and notes that “submitted URLs will be automatically shared with all other participating search engines” - one notification, many engines.

If the competitor is winning despite an internal defect. Microsoft’s December 2025 post on duplicate content is unusually specific about the AI mechanism: “LLMs group near-duplicate URLs into a single cluster and then choose one page to represent the set,” which “reduces the chances that your preferred page will be chosen as a grounding source.” That’s a live competitive opening: a competitor cannibalizing their own pages has a degraded grounding-selection problem you may not have.

What to ignore, and why ignoring it is a competitive decision

Google’s AI optimization guide has a mythbusting section, and for a competitor teardown it functions as a filter on bad advice you’ll otherwise be handed by vendors. Verbatim:

• On llms.txt and similar: “You don’t need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search (including its generative AI capabilities), as Google Search itself doesn’t use them.”
• On chunking: “There’s no requirement to break your content into tiny pieces for AI to better understand it.”
• On rewriting for AI: “You don’t need to write in a specific way just for generative AI search.”
• On mentions: “seeking inauthentic ‘mentions’ across the web isn’t as helpful as it might seem.”
• On scaled content: creating content per fan-out query “primarily to manipulate rankings or generative AI responses in Google Search violates Google’s scaled content abuse spam policy.”

If a competitor publishes near-identical pages for fan-out queries, the spam policy applies. Microsoft’s clustering description also says near-duplicate pages can be collapsed into one representative. That can affect which page gets grounded. I’d focus on content with something distinct to contribute.

Sources and References

• New AI Visibility Insights in Bing Webmaster Tools: Intents, Topics, Citation Share, Compare - Microsoft Bing Blog, June 2026. Source for Citation Share’s definition and the explicit “does not expose competitor domains” limitation, plus Intents and Topics classifications and Compare.
• Introducing Topics Portfolio (Beta) in AI Performance - Microsoft Bing Blog, October 13, 2026. Source for Sites Cited, Opportunity Queries, and beta caveats.
• Introducing AI Performance in Bing Webmaster Tools Public Preview - Microsoft Bing Blog, February 10, 2026. Metric definitions and their stated limits.
• Elevating the Role of Grounding on the AI Web - Jordi Ribas, Microsoft, February 2026.
• Does Duplicate Content Hurt SEO and AI Search Visibility? - Microsoft Bing Blog, December 19, 2025. Clustering behaviour in AI retrieval.
• Google’s Guide to Optimizing for Generative AI Features on Google Search - Google Search Central, last updated 2026-07-10 UTC. Non-commodity content, query fan-out, RAG grounding, mythbusting, Search Console eligibility and the Generative AI performance report.
• AI Features and Your Website - Google Search Central, last updated 2025-12-10 UTC. Technical requirements and control mechanisms. Conflicts with the 2026 guide on additional requirements.
• Guide to Preferred Sources in Google Search for Web Publishers - Google Search Central, last updated 2026-09-18 UTC. Preferred badge in AI Mode and AI Overviews; implementation.
• Creating Helpful, Reliable, People-First Content - Google Search Central, last updated 2026-10-05 UTC. Main content, user-generated contributions, Who/How/Why.
• Intro to Product Structured Data on Google - Google Search Central, last updated 2025-12-10 UTC. Product snippets vs merchant listings; review markup.
• Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact - arXiv:2605.14021, accepted to ACM IMC 2026, v2 October 1, 2026. 55,393 queries; ~30% of cited domains absent from first-page results; 11.0% of atomic claims unsupported by cited pages.
• The 50 Most-Cited Websites in Google AI Overviews (September 2026) - Ahrefs, updated September 2, 2026. Vendor research; US queries; mention share relative to top sources.
• The 50 Most-Cited Websites in ChatGPT (September 2026) - Ahrefs, updated September 2, 2026. Same methodology, different engine.
• Web search - OpenAI API documentation. sources vs inline citations, domain filtering, search modes, deprecation dates.
• Overview of OpenAI Crawlers - OpenAI. OAI-SearchBot, OAI-AdsBot, GPTBot, ChatGPT-User.
• Searching the web with ChatGPT - OpenAI Help Center. Query rewriting, search partners, Sources view, location, “Placement is not guaranteed.”
• Does Anthropic crawl data from the web, and how can site owners block the crawler? - Anthropic Help Center, April 7, 2026. ClaudeBot, Claude-SearchBot, Claude-User.
• Perplexity Crawlers - Perplexity documentation. PerplexityBot and Perplexity-User.
• Documentation - IndexNow.org. Submission protocol and key verification.
• What is the Model Context Protocol (MCP)? - Model Context Protocol. Standard for agent-to-system connections; relevant to integration-directory surfaces.