A Benchmark Guide to ChatGPT, Perplexity and AI Overviews Visibility

How do you compare brand visibility across three AI tools?

Compare brand visibility across artificial intelligence (AI) tools by asking the same set of questions under recorded conditions, then checking each answer against the same rules. Keep brand mentions, list order and source links as separate measures. This creates a benchmark: a repeatable test that lets a team compare like with like.

ChatGPT, Perplexity and Google AI Overviews offer different ways to find an answer. A chat without web search is not the same test as a search with source links. An AI Overview is a summary that may appear within Google Search, so first record whether one appeared.

The aim is to find clear next steps for your website and PR work. The test shows what happened within your chosen questions and settings. It does not measure the whole market or prove that a brand holds a fixed rank.

This guide explains how to run the test. The small number examples below are teaching examples, not results from a live brand study. No current brand league table is claimed.

Start with a business decision. A buyer looking for a new agency needs a different answer from a client checking a claim about an existing supplier. Keep those needs distinct so the team can see what each result means.

Which questions belong in a fair test?

A fair test uses questions that real buyers ask and keeps their wording stable from one run to the next. Gather them from sales calls, lost bids and requests for proof. Give each question a fixed code so it can be traced through every run.

Include some prompts that do not name your brand. A test built only around your name mainly shows how a tool responds to someone who already knows you. A buyer asking for a technology PR team in London creates a different test.

Build a small set around discovery, service scope and proof. For example, ask which firms suit the need, how an agency differs from a press release service, and what evidence a buyer should check. Those are separate decisions, even when the words overlap.

• Question: Save the fixed code and exact prompt, including any place or sector named.
• Purpose: State whether the buyer wants a shortlist, a comparison or proof of a claim.
• Settings: Record the date, time, language, country, account, plan and visible product mode.
• Session: Note whether the chat is new and whether search is active.
• Evidence: Save the full answer, source links and a screenshot under the question code.

Choose a repeat rule before you start. You might use a fresh chat for each question and keep the core set for three months. These are team choices for a clear test, not rules set by the platforms.

If a new service needs a new question, add a new version of the set. Do not quietly replace an old question and call the change a weekly gain. Keep the previous results with the wording that produced them.

Write each language version for its own readers. FL PR's guide to brand evidence across languages looks at this need for clear facts across markets. In a test, two natural prompts should target the same buyer decision without forcing a word-for-word translation.

What should you record when no AI Overview appears?

When no AI Overview appears, record that the feature was not seen and leave its brand rank empty. First check whether the answer area exists. Only then can you assess which brands it names.

Google's guide to AI features says that AI Overviews do not appear for every query. It also explains that responses and links can vary with the models and methods used. A normal search result is therefore not a substitute for a missing AI Overview.

Keep ads, standard search results and the AI summary in separate fields. A brand can appear in an ad but not in the summary. A source card can link to its site even when the answer text does not name it.

Consider a made-up run with nine questions. Suppose three produce an AI Overview and two of those three name the brand. Two out of three describes brand presence within the summaries, while three out of nine describes how often a summary appeared.

Two out of nine can also be shown as brand presence across all attempts. That measure combines two effects: whether a summary appears and whether it names the brand. Label each fraction so a reader knows what sits below the line.

• Seen: The AI summary appeared and its full answer was captured.
• Not seen: The search page loaded, but no AI summary appeared.
• Blocked: A sign-in step, usage limit or error stopped the check.
• Incomplete: Part of the answer or its sources could not be saved.

A blocked session is missing evidence, not a brand failure. Keep it in the log and state how many valid answers remain. This stops an access problem from being reported as a fall in brand performance.

How should mentions, list order and sources be scored?

Score a brand mention, its place in a list and its source links separately. A mention means the answer names the brand. A citation is a link offered as a source for part of the answer.

First look for the name in the answer text, not just in the source cards. Then check the kind of list the tool has produced. An alphabetical list or a set of examples does not state a ranked preference.

Save the raw order even when the list is unranked, but label it correctly. Compare firms that serve the same need. A software tool, a press release platform and an agency on a monthly fee may all be useful without being direct substitutes.

Open each source before giving it credit. Check whether it supports the claim beside the brand name and whether it describes the right firm. The correct domain can still lead to an old page, a different service or an unrelated person.

• Mentions: Count answers that name the brand and show the total number of valid answers.
• List order: Save the place, the list length and whether a ranked choice was stated.
• Sources: Store the exact URL, page title and date of the source check.
• Claim support: Note which sentence the linked page supports.
• Errors: Flag a wrong service, country, office, expert role or company name.

Average rank can hide missing mentions. If the brand appears in just two answers, an average based on those two says nothing about the other answers. Show absence counts beside the average and keep unranked lists out of a claimed ranking score.

Use raw counts when the sample is small. One mention in five valid answers is easier to assess than a large chart with a lone percentage. A move from one to two mentions needs further runs before it can be called a lasting trend.

Why might the platforms produce different results?

Different results need to be read alongside the product mode and the conditions of the test. OpenAI's web search guide explains that ChatGPT can search the web and provide links to sources. Record whether search was used instead of assuming that every answer used it.

Perplexity's product guide describes web search and cited sources as parts of its answer service. Source links still need a human check. A link does not prove that every sentence in the answer is right.

Keep fields for account state, plan, language, location, prior chat and visible mode. Do not infer a model version from a product name when it is not shown. Mark the version as unknown and retain the conditions you can see.

Keep the first valid response as well as later repeats. Choosing only the most flattering answer changes the test after seeing the result. Agree the number of repeats in advance and apply it to all brands and tools.

Look for three patterns: names shared across tools, names seen in only one tool, and facts that conflict. A repeated wrong service claim deserves a source check. A single different shortlist does not by itself show a fault.

The result can guide work on AI search visibility, often called GEO. FL PR's guide to what GEO work involves sets out the service scope. This test has a narrower job: it checks answers under known conditions rather than assigning a score to every part of that work.

How do the results become a useful work plan?

Turn each finding into a task with a named owner, a source and a due date. Start with facts that could lead a buyer to the wrong choice. A wrong service description is usually more urgent than a missing mention in one answer.

If the brand is absent, check that the question fits the service before writing new content. Next check whether the relevant page is clear and accessible. A useful page should answer the buyer's question and show the evidence needed to assess the claim.

Separate your own claims from third-party reporting. A firm can explain its services on its website, but proof of a result needs the original source and a clear account of its role. Paid placement and copies of the same press release should be labelled as such.

Give access issues to the technical team, missing answers to the content team and relevant press opportunities to the PR team. When a source contains an error, send its owner the correct facts and supporting proof. A request for correction is not proof that the source has changed.

Brand visibility and visits are also different measures. A source link may be seen without being clicked. FL PR's guide to measuring ChatGPT referral traffic covers the separate task of checking visits in site analytics.

Log the date of each site or source change, then repeat the same questions. If answers improve, check several runs before linking the change to your work. Products and other sources may have changed during the same period.

End the report with what was seen, what supports it and who owns the next step. Keep the raw evidence close enough for another person to check. That makes the benchmark useful for decisions, even when the tools disagree.

Frequently Asked Questions

These five checks help teams read an AI visibility test without overstating its results.