AI Overviews vs Perplexity vs ChatGPT: what each one actually cites

Ask three AI engines the same buying question and you get three answers that agree on the substance and disagree completely on who gets credit for it. That disagreement is the entire subject of this guide, because it is where the work is.

If you have not measured your own visibility yet, start with the 5-question AI visibility audit, which scores whether you get named at all. This guide is the layer under that one: not whether you appear, but what kind of source each engine reaches for, and therefore which change to your content is worth making.

What you'll have when you're done

• 15 logged answers: your 5 buying questions across Perplexity, ChatGPT and Google AI Overviews
• A domain tally showing which kind of source each engine favours in your niche
• A per-engine decision: the one content change that plausibly moves that engine, and the ones that do not
• A run log with dates, so run 2 and run 3 mean something
• One sentence you can say to a client or a founder about where you actually stand
• [SCREENSHOT: the three-tab citation log with two dated runs side by side]

Before you start

• Your 5 buying questions. The questions a customer types the week before they choose, none of them containing your brand name. If you do not have these written down, write them first. Everything downstream is only as good as these five lines.
• Perplexity, free tier. Citations are visible without an account.
• ChatGPT, free tier. Use a temporary chat so memory does not shape the answer.
• Google, in an incognito window. Note your country and city, because AI Overviews vary by location.
• A spreadsheet, and about 90 minutes.
• Optional: Claude or ChatGPT for the tallying step, which is dull and mechanical and worth handing over.

The three citation behaviours

Here is the shape of the thing before you run it, so you know what you are looking at.

Read the last row as the summary. AI Overviews is still substantially an SEO problem. Perplexity and ChatGPT are substantially a mentions-and-quotability problem. Teams that treat all three as one channel end up doing SEO work and wondering why Perplexity never changes.

A caveat that belongs here and not in a footnote: these behaviours are product decisions, and product decisions change. Everything below is a method for measuring the engines yourself, precisely so you are not relying on a description someone wrote six months ago, this one included.

Step 1: Fix the conditions before you run anything

Every result is worthless if run 2 is not comparable to run 1. Write the conditions at the top of the sheet and do not change them between runs.

RUN CONDITIONS
Date:            2026-08-29
Questions:       Q1-Q5, verbatim, unchanged since run 1
Perplexity:      logged out, new thread, default model
ChatGPT:         temporary chat, no memory, default model
AI Overviews:    incognito, google.com, location: Tokyo, JP
Order:           Perplexity, then ChatGPT, then Google

Three rules that decide whether the exercise is science or vibes:

1. Never edit a question between runs. Not for typos, not for phrasing. A changed question is a new question and the history resets.
2. Clean context every time. Memory, history and personalisation all shift answers, and they shift them in the direction of what you have already looked at, which is the exact bias you are trying to measure around.
3. Log the model version if the engine shows it. When a result jumps, this is usually why.

Step 2: Run all 15 and log the sources

For each question in each engine, capture three things: whether your brand is named anywhere in the answer, the ordered list of every domain cited or linked, and one line on what kind of answer it gave.

One row per source, not one row per answer. Long sheet, easy pivots:

source_type is the column that pays for the whole exercise. Use six values and nothing else: review, roundup, community, news, brand own, docs.

On AI Overviews specifically: run the query, screenshot the overview, then scroll and note the top 5 organic results too. You want the overlap number, because that is the single most decision-relevant fact about this engine in your niche. If 4 of the 5 cited pages are also in the top 5 organic, your AI Overviews strategy is your SEO strategy and you can stop treating it as a separate project.

Fifteen answers gives you somewhere between 40 and 80 source rows. Hand the tallying to a model:

Below are 15 AI answers to buying questions in one niche. Each is
under a heading like "Q1 / perplexity". Extract every domain cited
or linked, in order.

Output:
1. A table by ENGINE: engine, total citations, unique domains,
   and the count of each source_type (review, roundup, community,
   news, brand own, docs).
2. A table by DOMAIN: domain, times cited, which engines, which
   questions. Sort by times cited, descending.
3. The domains cited by all three engines. Label this "consensus set".
4. Domains cited by exactly one engine. Label this "engine-specific"
   and say which engine.
5. One sentence per engine: what kind of source it appears to prefer
   in this niche, based only on the counts above.

Do not include any domain that is not in the text below. Do not
infer, rank or editorialise beyond what the counts support.

ANSWERS:
[paste all 15]

Step 3: Read the result per engine

Now the tally becomes decisions. What you are looking for is not your score, it is the shape of each engine's preference.

If AI Overviews shows high overlap with organic results. Your lever is ranking for the question, plus making sure the passage at the top of that page directly answers it. Both parts matter: rank without a quotable passage often gets you the click-under rather than the citation. Write the answer in the first two sentences under a heading that is the question itself.

If Perplexity leans on reviews and community threads. Your own site is barely in this game. The lever is being named on the pages it already cites. Get listed on the review platform that appears most, ask five real customers to review you there, and answer the specific community threads that came up in your run, properly, as somebody who knows the subject. A thread cited today keeps being cited.

If ChatGPT names a short list of strong sources and never you. This usually reads as an entity problem rather than a content problem. It is naming things it is confident it can describe. The fix is upstream: one clear sentence about what you are, live in identical wording across your site, your schema and your profiles, and then third-party mentions that repeat it. The three-lever GEO playbook is the long version of that fix.

If the same domain appears across all three engines. That is the highest-value target on your list. One placement there registers three times.

Step 4: Why one run tells you almost nothing

This is the part most audits skip and it is the part that decides whether your numbers mean anything.

Run the same question twice in the same engine, an hour apart, in clean sessions. You will usually get an answer that agrees with itself on substance and cites a partly different set of sources. Nothing about you changed in that hour. The causes are all upstream: the retrieval pass surfaces a slightly different set, the index has moved, the model got updated, or there is plain sampling variation in what gets quoted.

The consequences are unforgiving:

• A single run cannot tell you a source is important. It tells you a source was reachable once.
• A single run cannot tell you a change worked. If your score goes from 3 out of 15 to 5 out of 15 after a month of effort, that gap is inside the noise you would see doing nothing.
• A single run absolutely cannot support a percentage. "You appear in 40% of AI answers" is a coin flip with a decimal point on it.

What does survive:

1. Repeat sources. Domains cited in run 1 and run 2 and run 3. That is your real competitive set, and it stabilises fast, usually by the third run.
2. The source-type mix per engine. The exact domains shuffle. The proportion of reviews to community threads to brand sites is much steadier, and it is what you are actually planning against.
3. Direction over three runs. Three consecutive runs pointing the same way is a signal. Two is a line drawn through two points.

Step 5: Log runs so you get a trend

Three tabs, one sheet.

Tab 1: runs. One row per source, the long table from step 2, appended forever. Never overwrite a run.

Tab 2: summary. One row per run date per engine:

Tab 3: targets. Every domain seen twice or more, with the runs it appeared in, its source type, a realistic path to being mentioned there, and the date you acted. This tab is the actual work list, and it is the only one anybody else needs to see.

Three numbers to read each month, and no others:

1. Named count out of 15. Blunt, comparable, honest.
2. Engine spread. 5 out of 15 concentrated in one engine is a completely different position from 5 spread across three. The first is an engine-specific quirk, the second is real presence.
3. Repeat-source count. How many domains carried over from last run. Rising means your target list is firming up, and the work gets more efficient every month.

What not to track: how many pages you published, how many pitches you sent, whether rankings moved. Those are inputs, and inputs are what people report when the output has not moved.

Where this is going

The method above works, and it is 30 minutes a month forever, done by hand, by a person who could be doing something else. That is exactly the shape of a job that should be a scheduled job: fixed questions, fixed engines, fixed sheet, run on a timer, with the history kept and the trend drawn for you.

That is what I am building CiteScore for. Same 5 questions, same 3 engines, run on a schedule, sources tallied and typed automatically, and the only thing surfaced is what changed since last time. The reason it exists is the whole of step 4: the useful signal in this data lives across runs, and nobody reliably keeps a manual log going past month two.

Run it by hand first anyway. The manual runs are where you learn what a good source looks like in your niche, and that judgment is the part no dashboard hands you.