How to Check What ChatGPT and Perplexity Say About Your Business
- 8 hours ago
- 7 min read
Check what ChatGPT and Perplexity say about your business by running a small prompt library in a fresh session, more than once, and scoring named versus cited versus wrong. One screenshot is noise. Citation decisions swung 9% to 28% between identical runs at temperature zero (Kirsten et al., 2026, in arXiv:2607.14035).
An AI visibility check is a repeatable prompt test, not a casual chat in a logged-in thread. The leftover hub mentioned a five-minute glance. This page is the shareable version.
Key Takeaways
Score named and cited as two columns. A shop can be linked and never named, or named and never linked.
Run each prompt at least 3 times in a fresh, logged-out session. The 2026 critical survey recommends 7 to 8 repetitions as a baseline (arXiv:2607.14035).
ChatGPT skipped web search in 57.8% of repetitions of the same prompt in that survey. A no-search answer is data, not a void.
ChatGPT recognized a named product on 99.4% of direct asks and surfaced it on 3.32% of unbranded discovery queries. Perplexity sat at 94.3% versus 8.29% (arXiv:2607.14035).
BrightLocal found 45% of 1,002 US adults had asked AI for a local recommendation in the past year (BrightLocal). The unbranded city prompt is the one those people type.
The Short Version
1. Copy the 12 prompts. 2. Open a fresh ChatGPT session and a fresh Perplexity session, logged out, no memory. 3. Run each prompt 3 times. 4. Fill the split sheet: named, cited, sources, wrong facts. 5. Fix the pages and listings the answers already used.
Field | What you write down |
Engine | ChatGPT, Perplexity, Gemini, Copilot |
Prompt | The exact sentence you pasted |
Named? | Yes or no, in the prose |
Cited? | Yes or no, as a source link |
Sources | Every URL or site the answer showed |
Wrong? | Hours, address, services, or city errors, verbatim |
In This Article
The Split Sheet
The split sheet is a five-column log you fill on every run: engine, prompt, named, cited, and sources. Mentions and citations diverge. Vendor cuts have put Gemini mention-to-citation overlap as low as 30%.
RedFork has helped more than 800 businesses grow and built 500+ websites since 2014. The Partners who treat one branded ChatGPT tab as a report come back surprised the next week. The Partners who keep a sheet see the swing and stop overreacting.
Print the sheet or copy it into a notebook. Date every row. Canonicalize URLs before you count. example.com/page and www.example.com/page/ are one citation, not two. That rule comes from the same 2026 survey that called single runs noise.
Prompt Library for a Local Shop
The prompt library is 12 sentences a neighbor would type, split into branded checks, unbranded city prompts, and problem-led asks. Rewrite the brackets. Do not add your brand to the unbranded set.
Branded (3). These test recognition and fact accuracy, not recommendation.
What do you know about [Business Name] in [City]?
Is [Business Name] a good [trade] in [City]? What are the hours and the address?
How does [Business Name] compare to other [trade] options in [City]?
Unbranded city (6). These are the product.
Who should I call for [job] in [City]?
Best [trade] near [neighborhood or city]?
I need [service] this week in [City]. Who do you recommend?
Which [trade] in [City] is open Saturday and takes walk-ins?
Affordable [service] in [City] for a small job?
Who does [specific job] in [City] and what should I budget?
Problem-led (3). These test whether the model reaches for you without a category noun.
My [thing] broke in [City]. What do I do, and who can help today?
I just moved to [City] and need a regular [trade]. Where should I start?
How do I tell a good [trade] from a bad one in [City]?
Cross-engine overlap is low. Domain overlap between Bing Chat and Perplexity sat at 26% in one survey cut. Source-level Jaccard overlap across four engines sat at 0.34 to 0.42 (arXiv:2607.14035). One engine's row tells you nothing about the next.
How Do You Run ChatGPT Without Contaminating the Test?
You run ChatGPT in a fresh, logged-out session with memory off and no custom instructions, then paste one prompt and stop. Follow-ups train the thread. That is a conversation, not a test.
OpenAI states ChatGPT Search may share disassociated queries with third-party providers such as Bing (OpenAI, ChatGPT Search). Note whether the run actually searched. A 57.8% skip rate means more than half of identical prompts never touched the live web. Record that as "search: no." Do not score a memory answer as a retrieval miss.
Keep the location consistent. If you live in one city and serve another, say the city in the prompt. A "near me" with no city is a lottery. Same day of month, same window, if you repeat the audit. Personalization is contamination.
How to Read a Perplexity Answer
A Perplexity answer is a short list plus visible sources. Read the prose for the name, then read the source list for the URL. Those two lines are the named and cited columns.
Perplexity always retrieves. That makes it the easiest engine to score, and the most likely to show you the exact page it trusted. If your shop is named and the sources are Yelp, Reddit, and a competitor's blog, the fix is not another homepage paragraph. If your URL is cited and the prose names someone else, you won the footnote and lost the recommendation.
Perplexity's unbranded discovery rate in the 2026 survey was 8.29%, against 94.3% branded recognition (arXiv:2607.14035). A clean branded answer and a silent city prompt is the common local pattern. Write both rows down. Do not average them into one "AI score."
Are Named and Cited the Same Result?
Named and cited are two different outcomes on the same ChatGPT or Perplexity answer about a shop. Named means the shop appears in the sentence a customer reads. Cited means a source link points at you. Wrong means the answer invented or stale-dated a fact.
Score them apart. A mention without a citation is still a recommendation. A citation without a mention is a footnote. Wrong hours, a closed location, or a service you do not offer is a trust problem even if the name is right. BrightLocal found 88% of AI users fact-check a recommendation against a real review or a source (BrightLocal). The customer will see the error.
Copy the wrong sentence verbatim. That is the rewrite brief for the page or the listing that fed it. Do not argue with the model in the same thread. Fix the source, then re-run next month.
How Many Runs Before You Trust a Screenshot?
You should not trust one ChatGPT or Perplexity screenshot as a visibility score for your business. Citation decisions swung 9% to 28% between identical runs at temperature zero (Kirsten et al., 2026). Three runs is the floor for a shop-owner audit. Seven to 8 is the research baseline.
Look at the pattern, not the hero run. If you are named on 1 of 3 ChatGPT runs and 2 of 3 Perplexity runs, you have a rate, not a verdict. If run 1 names you and run 4 names two others, both rows stay on the sheet. About 45% of cited sources churn per model update in the engine evidence base. Measure monthly. React quarterly.
Paid tools can automate the sheet. Manual is enough to start. Rankscale, Otterly, and Rankshift exist if you later need share of voice across Partners. The sheet is the habit. The tool is optional.
What to Do With a Bad Week of Answers
A bad week of answers is a source list, not a reason to buy a third acronym. Read the sources the engines already used, then make those pages and listings tell the same story.
If branded prompts fail, the entity is thin. About page, matching listings, and a consistent name-address-phone string come first. If branded prompts pass and unbranded city prompts fail, you have recognition without recommendation. That is the 99.4% versus 3.32% gap. If the sources are directories, go there. If the sources are a competitor's city page, write the extractable sentence you are missing.
Express websites on curated Wix templates cover a one-location site plus Ads Only marketing. Pro custom Wix work with the full marketing mix is the better fit when the page set and content need to change. Unserious Marketing will not replace a split sheet. If you want the audit scoped, GET PRICING. We will not promise the next run names you.
How We Researched This
On 28 August 2026 we ran WebSearch on "how to check what ChatGPT and Perplexity say about your business" and the close PAA variants, then compared DIY audit pages against the 2026 critical survey's reproducibility rules, BrightLocal LCRS 2026, OpenAI's ChatGPT Search note, and dossier Section 7. Live RedFork URLs in this article were curl-checked the same day and returned HTTP 200. The prompt library is original to this page and sized for a local shop, not a 150-prompt enterprise study. No affiliate relationships.
Frequently Asked Questions
Can I run this inside my usual ChatGPT account?
You can, and you will contaminate it. Memory, custom instructions, and prior turns change the answer. Use a fresh session. Logged out is cleaner.
What does a branded prompt prove?
A branded prompt proves recognition and fact accuracy. It does not prove a neighbor will hear your name. Use it to catch wrong hours. Use the unbranded set to score recommendation.
Should I include Gemini and Copilot?
Yes, if you have time. Cross-engine overlap is low. Copilot leans on Bing. Gemini leans on Google. A ChatGPT-only sheet is a ChatGPT report.
How often should I repeat the audit?
Monthly is enough for a one-location shop. Same week, same window, same prompt set. Do not refresh the score after every model headline.
Do I need a paid AI visibility tool?
No. The sheet is free. Add a tool when you need share of voice across many prompts or many locations. Manual first.
If the Sheet Is Full, Fix the Sources It Named
Keep the library. Date the rows. Change the pages and listings the answers already trusted.
If you want that work scoped on a Wix site, GET PRICING. Express if you need a template and ads. Pro if you need the full mix. More background sits on the RedFork Learning Center. We will not treat one screenshot as a launch date.
Latest Updates
30 August 2026: Drafted from live SERP recon, arXiv:2607.14035 reproducibility findings, BrightLocal LCRS 2026, and dossier Section 7 credentials.
References
Related Reading
This article is educational. RedFork does not guarantee rankings, citations, traffic, or revenue. Statistics were current as of 30 August 2026 and should be re-checked on refresh.
Comments