How to track AI brand mentions across ChatGPT and Perplexity

What counts as an AI brand mention across both platforms?

A mention counts when the generated answer explicitly identifies the intended brand or an unambiguous approved alias. A source link, a recommendation, and a factual description remain separate observations.

Begin with a written identity rule. Record the trading name, product names, common abbreviations, previous names, and domains that genuinely belong to the company.

Mark ambiguous abbreviations for human review rather than counting every text match. A short brand name that is also an ordinary word needs contextual confirmation, such as a product description or a matching company domain.

Keep answer text and source lists distinct. A brand may appear only in the title of a cited article, while the generated prose discusses an unrelated vendor.

That is a source-list appearance, not automatically an answer-body mention. Conversely, the answer may recommend your product while citing an independent review rather than your website. Both observations can matter, but they answer different questions.

Define recommendation fit against the buyer constraints in the prompt.

If a hypothetical SaaS buyer requires an available integration and the answer recommends your product despite an unsupported integration, record inclusion and a fit failure.

A positive tone does not repair an inaccurate recommendation. Preserve the exact statement so that a product owner can evaluate the same evidence.

This separation makes the paired comparison useful. ChatGPT might name your product without citing your domain, while Perplexity cites your tutorial without recommending the product.

Neither result should overwrite the other in an aggregate score. Your ledger should explain which kind of visibility occurred and whether a buyer could rely on the answer.

How should you build a matched prompt set?

Use the same buyer intent and constraints on both platforms, with branded and unbranded prompts stored in separate groups. Freeze the wording for each comparison period.

Start from sales discovery notes, support questions, and actual evaluation criteria rather than a list of broad category keywords. Choose questions that could plausibly affect a buying decision.

For a fictional reporting product, one unbranded prompt could state, “Compare subscription reporting tools for a finance lead who needs documented billing integrations and explain each option’s limitations.” This requests a real evaluation without planting your brand.

A branded diagnostic prompt could state, “Assess whether Example Analytics supports the billing integration described on its current public documentation and cite the relevant sources.” Treat this as an accuracy test.

Because the question supplies the brand, its appearance does not demonstrate spontaneous discovery. Keeping that distinction explicit prevents a monitoring programme from rewarding itself for repeated brand insertion.

Create matched groups for category discovery, a narrow use case, an alternative evaluation, and a purchase constraint. Keep the buying stage attached to every prompt.

A category explanation and a procurement comparison should not carry the same interpretation simply because both contain your product name. Decide which groups matter most before inspecting the answers.

Use identical wording where practical, but do not pretend identical wording makes every setting identical. Store the prompt version, language, stated geography when relevant, and any supplied links.

If a question changes midway through a reporting period, create a new version rather than silently replacing the old one. Preserve the old version for a comparable historical baseline.

Which ChatGPT and Perplexity conditions need recording?

Record the interface, visible mode, search evidence, date, and conversation context for every run. A consumer interface and an API workflow are different measurement surfaces.

OpenAI explains that ChatGPT may search automatically when current information would help and that a user can also initiate search. Its managed-workspace documentation describes controls that can affect search availability.

Record whether you observed search activity or source links instead of assuming every answer used the live web. An answer without observable search should carry that uncertainty in the record.

For Perplexity, record the product surface and the mode actually displayed in your account, together with any source restrictions you apply.

Avoid hard-coding a universal menu label into the monitoring procedure because interfaces can change. If you deliberately limit sources or supply a specific page, label the result as a constrained test rather than open discovery.

Begin matched discovery tests in fresh conversations and avoid uploading proprietary brand material into those sessions. Previous turns can supply a brand, a competitor, or a preferred evaluation criterion that changes the task.

Record account and workspace context at an appropriate privacy level. Keep confidential account identifiers outside any broadly shared marketing report.

Run both sides within a defined observation window and preserve timestamps. If a website release occurs between the paired runs, flag the pair rather than attributing the difference entirely to the platforms.

Automation can help capture consistent fields, but an API model configured by a vendor should not be relabelled as an observation from the consumer application.

Sources: OpenAI explains web search in ChatGPT. · OpenAI documents search controls for managed workspaces.

What should each row in your evidence ledger contain?

Give each generated answer its own immutable run identifier and preserve the original response before extracting metrics. Store enough context to reproduce the review, not just the final score.

A practical record contains a prompt identifier, prompt version, platform, interface, visible mode, timestamp, conversation context, search evidence, run status, answer text, and captured source URLs.

Add separate fields for brand identity, answer-body inclusion, recommendation fit, factual accuracy, and citation support.

These fields describe observations rather than undocumented assumptions about how the platform generated them.

Save screenshots or exports when available, subject to your organisation’s privacy and retention rules.

Keep a readable transcript alongside the structured row because citation markers and answer formatting can matter during review.

If a citation opens through a redirect, preserve the original reference and the destination you inspected. Do not strip information that a later reviewer needs to trace the claim.

Assign operational statuses such as completed, unavailable, interrupted, or awaiting review. A failed page load is not evidence that your brand was absent from an answer.

Likewise, an answer that stops midway should not be quietly scored against a completed response on the other platform. Include exclusions and their reasons in the report so that the sample cannot appear cleaner than it was.

For extraction, start with conservative text matching and require human review for ambiguous names, implied recommendations, and factual disputes. Keep the original extracted value and the reviewed value if a correction is made.

A small, well-audited ledger is more defensible than a large collection of unsupported positive or negative labels.

How do you check whether a mention has meaningful citation support?

Open the cited page and compare its actual content with the claim attached to the citation. A valid link does not establish that the answer interpreted the page correctly.

Inspect the specific capability, limitation, date, or comparison the answer attributes to the source. A generic homepage may identify the company while saying nothing about the integration being discussed.

Record whether the page supports identity only, supports the particular capability, contradicts the answer, or cannot be evaluated. Keep unsupported and inaccessible cases distinct.

For a fictional example, suppose ChatGPT says a reporting product supports a native data warehouse connector and cites a guide describing a manual export.

The brand mention is real, but the capability statement is unsupported by that passage.

If Perplexity instead describes the manual workflow accurately, the comparison reveals an accuracy difference, not simply a difference in brand visibility.

Read independent pages with the same scrutiny as your own documentation. An old comparison may describe a discontinued plan, and a directory entry may reflect a previous company name.

Preserve the publication or update date when visible and note the relevant scope. Do not treat third-party repetition as confirmation when the underlying statement cannot be checked.

Citation auditing also prevents a misleading domain count. Several links to your site might all support general background rather than a buying recommendation.

Record the claim-source relationship first and aggregate domains second.

This produces a useful editorial backlog because you can see which buyer statements lack accessible, current evidence and which answers simply misread evidence that already exists.

Sources: OpenAI explains web search in ChatGPT.

How can you compare the results without creating misleading rankings?

Calculate platform-specific rates from completed eligible runs and show paired disagreements separately. Label every denominator and keep prompt categories visible.

Define answer-body mention rate as completed eligible answers containing a verified brand mention divided by completed eligible answers in the same group.

For unbranded discovery, exclude branded diagnostic prompts from that denominator. Define citation presence separately, and make clear whether it means any cited company page or a citation that supports the relevant statement.

These are recommended operational definitions, not official platform metrics.

For each matched pair, assign one of four inclusion outcomes: both platforms mention the brand, only ChatGPT mentions it, only Perplexity mentions it, or neither mentions it. Then review accuracy and fit inside those groups.

This view exposes asymmetric outcomes that disappear when the two platforms are averaged into one number.

Repeat the fixed prompt set on a declared cadence and keep each response as a separate observation. Multiple runs can reveal observed variability, but they do not establish a representative sample of all users.

Avoid publishing confidence claims without a defensible statistical design. Report the number of prompts, runs, exclusions, and periods so that the reader can judge the scope.

Use a stable core set for trend reporting and a separate exploratory set for new buying questions. If the exploratory set grows, report it independently until you establish a baseline.

Otherwise, a change in question mix can look like a change in platform performance. Do not use answer order as a universal ranking position when the response is a narrative rather than a ranked list.

What should you investigate when the two platforms disagree?

Use disagreement to identify an evidence or access question, not to infer a hidden ranking formula. Start with the actual cited pages and the conditions of the paired runs.

First check whether the pair is genuinely comparable. A search-enabled response and an answer with no observable search may be using different evidence contexts.

A follow-up conversation and a fresh session also answer different tasks. Resolve these setup differences before deciding that your content is performing differently across platforms.

Next inspect public access. OpenAI distinguishes OAI-SearchBot, which supports search, from GPTBot, which concerns potential training collection.

Perplexity documents PerplexityBot for search visibility and Perplexity-User for user-requested page visits. Review the relevant robots policies, firewall handling, and actual delivery logs with your web owner.

A policy decision about training should not be mistaken for a search visibility diagnosis.

An allowed crawler is not proof that a page was retrieved for a particular answer. Equally, a missing citation is not proof of blocking.

Use server evidence to investigate access, and use the saved answer to establish what was displayed. Avoid causal language unless the observation supports it.

A corrected firewall rule and a later citation are related events, not an automatic controlled experiment.

Finally, compare the evidence available for the disputed buyer requirement. If your documentation leaves an integration ambiguous, publish a precise answer with limitations and an update owner.

If the current page is already clear, monitor whether the error persists and consider the platform’s available feedback mechanisms. Prioritise commercially consequential inaccuracies over cosmetic differences in wording.

Maintain a change log for your own monitoring process as well as your website.

A different account, newly selected mode, revised identity matcher, or altered exclusion rule can change the reported result without any underlying improvement in buyer visibility.

Annotate these changes on trend charts and, where possible, rerun the previous procedure alongside the new one before comparing periods.

Sources: OpenAI documents its crawlers and user agents. · Perplexity documents its crawlers and user-requested fetcher.

How should marketing, product, and analytics use the findings?

Turn each consequential finding into an owned action with a supporting run identifier and a retest plan. Keep monitoring, factual correction, and commercial attribution separate.

A weekly review can sort findings into access barriers, missing public evidence, inaccurate descriptions, unsuitable recommendations, and stable accurate appearances.

Assign the web owner to delivery problems, the product owner to capability verification, and marketing to evidence presentation.

The point is an accountable investigation rather than a growing dashboard of unexplained red indicators.

Write each action in a form another person can verify.

For example, state that the integration page must distinguish a native connector from a supported export workflow, identify the disputed answer, and define the public source that will resolve the ambiguity.

Preserve the before-and-after page state and the publication date. Retest the original prompt version after the change without claiming that one changed answer proves causality.

Connect the monitoring record to commercial evidence cautiously. Referral sessions, conversion events, and prospect-reported discovery can add useful context, but an observed mention does not prove a lead saw that exact answer.

Ask prospects how they researched the category and preserve their own description. Keep inferred influence clearly separate from observed visits or reported exposure.

AEO Operators can help define the paired protocol, review citation quality, and turn findings into a content and access backlog. We are a service agency rather than monitoring software.

Our free AI Search Grader is a one-time report on how ChatGPT, Perplexity, and Gemini describe a brand; it can inform an initial diagnosis but does not replace repeated matched observations.

Set an escalation rule before the first reporting cycle. An inaccurate security statement or unsupported contractual commitment deserves faster review than a missed appearance in a broad educational answer.

Keep the severity rationale attached to the evidence so that repeated minor issues do not crowd out a material buyer risk.

The resulting priority list should reflect commercial relevance and factual consequence, not merely the number of times a brand name appears.

Frequently asked questions