A useful ChatGPT–Perplexity citation test tells you whether buyers find your business, not whether an assistant will link to your website when asked. Most tests make that mistake before the first prompt runs: they pick questions the brand already wins, then call the result visibility.

I want a different result: a fixed panel of 15 to 25 buyer questions, run consistently in both engines, that separates being cited as a source from being recommended as a supplier. That distinction helps decide whether the next job is a better page, better third-party proof, or nothing at all.

If the service measuring your visibility also chooses the questions, this is the part to get right first. Otherwise, you are paying someone to grade their own homework.

What should a citation test actually measure?

A supplier-list check mixes together three events that need different work to fix. A mention means your brand is named, a citation means your page is linked as a source, and a recommendation means you are presented as a good pick.

Those are not interchangeable wins. One study found ChatGPT linked the recommended brands’ own sites in 70% of cases, while Gemini did it in 27% and AI Overviews in 21%. An engine can recommend you using someone else’s page as evidence. It can also cite your explanation without recommending you as a supplier.

Asking an assistant to cite your website does not establish buyer visibility. Branded prompts test accuracy, while unbranded category, problem, and comparison prompts test whether you get found without help.

Keep those jobs separate. A branded answer is an accuracy check. An unbranded recommendation is a discovery result. Neither should borrow credit from the other.

Step 1: Which buying decisions belong in the panel?

Say you sell commercial HVAC maintenance. I would start with four decisions, not a keyword export:

  1. Discovery: “How do I cut downtime in a 50,000 sq ft warehouse?”
  2. Comparison: “Preventive maintenance contract vs. on-call repair for warehouses.”
  3. Qualification: “What should a food-grade warehouse require from an HVAC contractor?”
  4. Purchase constraints: “Typical cost and response time for warehouse HVAC contracts in Chicago.”

Build a panel of 15 to 25 prompts around those decisions, with branded checks and explicit source requests kept in separate diagnostic groups. If a discovery prompt has no plausible connection to a buying decision, cut it.

Hand-picked brainstorm lists drift toward what the team hopes is true. I have done it myself: twenty clever questions where our brand inevitably appeared. Build the panel from evidence you did not invent instead: site search queries, server-log questions, AI referrals that actually landed, and the exact language from sales calls.

Use the buyer’s wording, even when it is less polished than the phrase in your positioning deck. Then freeze it. Rewriting a question until your brand appears is not improving the test.

Five real customer questions beat fifty invented ones. Skip manufactured prompt volume.

Step 2: Which prompts should stay separate?

Put every prompt into one of four groups. Comparison prompts belong in the unbranded results only when they do not name your brand.

GroupWhat it asksWhat I would score
Unbranded discoveryA category or problem question with no brand nameUnaided mentions, citations, and supplier recommendations
ComparisonOptions, prices, or tradeoffs within a categoryShortlist visibility; keep brand-naming comparisons separate from unaided discovery
BrandedA question that names your business directlyAccuracy of services, prices, locations, and constraints
Explicit source requestsA request such as “cite your sources for warehouse HVAC maintenance costs”Citability under an explicit request, not ordinary buyer visibility

What the deck calls prompt coverage, I call mixing the exam with the answer key. A prompt that says “cite my site” tells you little about the warehouse manager who types a plain English problem into Perplexity.

Never blend branded checks or explicit source requests into your discovery score. Keep them in the sheet because they answer useful questions. Just do not let an easy diagnostic win disguise a discovery miss.

Step 3: How do I make the engine comparison fair?

Keep the buyer question word for word the same and record the conditions. I would log these before scoring anything:

  • Engine and model.
  • Date.
  • Web-search mode.
  • Location and account context, where visible.

The mode matters because Perplexity generally searches the live web and supplies numbered citations, while ChatGPT’s source links depend on search use and the answer being generated. Run ChatGPT with search on for the search comparison. If you also test it with search off, keep that as a separate condition rather than folding both into one ChatGPT score.

Otherwise, the sheet mixes what an engine retrieves today with what it answers without live search. That is not a clean comparison.

Next, repeat the runs. In one 50-prompt test repeated 23 minutes apart, the identical brand set repeated only 60% of the time when a brand was named, and only 41 of 200 pairs cited exactly the same domains.

My run protocol is simple:

  1. Open a fresh chat, using temporary chat where available.
  2. Keep location, account state, and the selected model consistent within each engine’s test. Use logged-out access where possible.
  3. Submit the exact saved prompt without a follow-up that steers the answer.
  4. Save the response and its links.
  5. Repeat each prompt three times on the agreed schedule.

Expect the engines to disagree. One comparison found only about 11% of domains were cited by both ChatGPT and Perplexity, with different source preferences across the engines. The point is to preserve those differences, not average them out.

A fixed panel repeated under recorded conditions is measurement. A single favorable answer is an anecdote.

Step 4: What do I save, and how do I score it?

A visibility percentage alone gives your strategist very little to fix. Save the whole answer, not just the flattering sentence.

Use one row per run and log the engine, date, mode, brand mention, list position, cited URLs, competitors named, and accuracy. Add the prompt ID, full answer text, suppliers recommended, and the page on your site that should answer the question.

Mark answers without citations as no citation. Mark failed runs as failed. Do not delete either, and do not count a failed run as an answer that omitted you. A blank cell should not have to explain itself later.

For completed unbranded answers, calculate these separately for each engine and search condition:

  • Mention rate: answers naming your brand divided by answers checked.
  • Own-domain citation rate: answers linking your domain divided by answers checked.
  • Recommendation rate: answers presenting you as a suitable supplier divided by answers checked.
  • Share of voice: your presence alongside the named competitors, using the same counting rule throughout.

Also distinguish a citation to your page from a citation to a third-party page that mentions you. Both matter. Only one is an own-domain citation.

This is where the test starts earning its keep. A page cited for a useful explanation is not automatically a supplier endorsement. A supplier recommendation supported by a third-party review is not a failure just because your domain receives no link.

Read mentions, citations, and recommendations together, but never collapse them into one win. When the numbers diverge, inspect the saved answers before assigning the work.

Step 5: What job does each repeated miss create?

I want every repeated gap to produce a task, not another chart. Start with four possible jobs:

  • Missing buyer answer: If no page answers the decision plainly, create or rebuild the relevant page around costs, constraints, and comparisons.
  • Wrong business information: If the answer repeats old hours, prices, or service areas, correct the source it trusted. Check your site, then the profiles and listings it cited.
  • Inaccessible content: If the page exists but cannot be fetched, check crawlability and indexing before rewriting it.
  • Missing outside proof: If the answers rely on third-party evidence you lack, identify the sources trusted for that question and the evidence gap to address.

Do not assign those jobs from a percentage alone. A high mention rate and low own-domain citation rate tells you to inspect the gap; it does not tell you why it exists. Look at what the engine actually cites. The answer may point toward clearer definitions, costs, comparisons, or tables on your site. It may instead point toward outside proof.

Access needs its own check. To make pages available to these search crawlers, allow OAI-SearchBot for ChatGPT Search and PerplexityBot for Perplexity through robots.txt and relevant CDN rules; GPTBot training access can be handled separately.

But crawler access is not a citation, and a citation is not a recommendation. A fetch in a server log shows that a page could be read. It does not show that a buyer saw your business recommended.

The useful output is an assigned queue: which page changes, which source needs correcting, and which evidence gap needs outside coverage. The 30-prompt test shows how buyer prompts, repeated runs, and citation counts fit on one sheet.

Step 6: How do I keep the baseline honest?

Freeze a core set of 15 to 20 prompts within the panel and version anything you add. When you change wording to chase a citation, you lose the ability to say whether visibility improved or you simply asked an easier question.

I keep one tab locked and one for experiments. New buyer language goes into the experiment tab for a month. Only prompts that still match real sales conversations move into the locked set, with a date stamp.

Keep engine and search conditions separate when comparing results. For each prompt, record mentions, own-domain citations, and recommendations as 0 of 3, 1 of 3, 2 of 3, or 3 of 3 for the review period. Then inspect the prompts stuck at zero and the ones that differ sharply between engines.

A prompt that cites you 3 of 3 times in Perplexity and 0 of 3 in ChatGPT is a map of different evidence choices, not a reason to rewrite the question. The saved answers tell you where to investigate.

Keep fetch logs in their lane, too. I used to forward them to clients as good news. I stopped when I realized buyers never see a log line.

A crawler fetch tells you the door was open. Repeated answer captures tell you whether your page was cited or your business recommended. Measure the event you actually care about.

What should I ask my groas strategist to deliver?

Send your strategist the sheet, not a score. I would ask for four things back:

  1. A shared prompt register: the locked core, separate diagnostic groups, and dated experiments.
  2. Baseline captures: full answers and cited URLs for every run, with engine and search conditions recorded.
  3. An assigned action queue: one job per repeated miss, with an owner and a named page or source.
  4. A review date four weeks out: the same panel, run again under the recorded conditions.

No promise of rankings, citations, or leads belongs in that handoff. No service controls what ChatGPT or Perplexity retrieves on a given day.

What groas controls is the execution after the test: the page built, the wrong fact corrected at its source, the thin answer rewritten so a model can quote it. The work is the product, not the report. A prettier score does not excuse a panel chosen because it already produces citations.

What should I do today?

Start smaller than the full panel, but keep the same discipline:

  1. Pull ten real buyer questions from sales calls, site search, or support tickets.
  2. Turn them into twelve unbranded prompts in the buyers’ words.
  3. Run each prompt three times in ChatGPT with search on and three times in Perplexity, using fresh chats and logged-out access where possible.
  4. Save every answer and its links. Mark mentions, own-domain citations, and supplier recommendations separately.

If those twelve prompts produce two mentions and zero citations, do not respond by inventing more favorable prompts. Read the answers. Check whether the missing work is a quotable page, outside proof, or something else in the evidence you saved.

The four-week audit method gives you a longer follow-up for checking fetches and ordering content gaps. Today, pull the ten buyer questions first. Hand that baseline to your groas strategist and build the work queue from what buyers actually ask.