Ask Gemini and ChatGPT for the best payroll software for a restaurant group, and your brand can make one shortlist while missing the other. The assistants do not read the same web, so visibility in one does not guarantee visibility in the other.
One study ran 38,966 unbranded buyer questions on both assistants and found they agreed on consistently recommended brands only 42% of the time. That is not a reason to chase every answer. It is a reason to stop treating “AI visibility” as one score.
Gemini leans on Google’s search index when it grounds answers in search. ChatGPT combines training data with its own search and retrieval paths. The same buyer question can therefore surface different sources, then different brands. Everything else follows from that distinction: measure the assistants separately, diagnose the sources behind the gap, and fix the one your buyers actually use first.
Start with one buyer, not two screenshots
Say you sell payroll software to a 20-person restaurant group with hourly staff and tip reporting. I would start with an unbranded question: best payroll software for restaurants with hourly staff.
One answer might include Rippling, Gusto and Toast Payroll. Another might include Gusto, ADP Run and OnPay. Those are illustrative shortlists, not a prediction of what either assistant will return today. The point is that the buyer’s need stays constant while the recommendations change.
There is another moving part: even the same assistant does not reliably give the same list twice. In one test of about 3,000 runs, the chance of getting the identical brand list twice was less than 1 in 100, and about 1 in 1,000 for the same order. The strongest brands still appeared in 60 to 90% of answers.
The model can select different candidates across runs, so one screenshot cannot tell you whether your brand has a persistent visibility problem. It tells you what happened once. Useful for a slide, less useful for a decision.
Measure how often you appear across repeated answers, not whether you appeared on Tuesday. That gives us something worth diagnosing.
Gemini: start with the sources Google surfaces
When Gemini needs fresh information, it can use Google Search to ground an answer. Google documents this as grounding responses in its search engine results.
For our payroll company, that makes the Google-visible source landscape a sensible first inspection point. Search for restaurant payroll with tip reporting. Which roundups, reviews, discussions and product pages appear? Which brands do those pages name?
If those sources repeatedly recommend competitors and omit you, Gemini has fewer search-grounded routes to your brand. That is different from saying a page-two ranking makes you invisible. A ranking is a clue, not a complete explanation of an answer.
The comparison discussed by ORR Consulting found ChatGPT citing brands’ own sites in 53% of answers versus 13% for Gemini. In that comparison, Gemini leaned more heavily on outside sources such as roundups, Reddit and trade press.
For the payroll example, I would inspect:
- Review pages that describe restaurant-specific payroll needs.
- Discussions where owners name products and explain why they use them.
- Roundups that explicitly cover hourly staff, tip reporting or tip pooling.
- Your own pages answering those same needs.
A polished features page cannot fill every missing third-party mention. If Gemini keeps citing pages that leave you out, work on that source gap—not just your homepage copy.
ChatGPT: inspect retrieval access and your own pages
ChatGPT can build an answer from training data, live search, licensed publisher content and the conversation itself. For live search, it can rewrite a question into narrower queries and assemble an answer from the pages it retrieves.
That gives our payroll company another diagnostic path. Instead of looking only at the broad category query, I would inspect the specific questions behind it: tip pooling, hourly payroll, pricing and support for restaurant workflows.
Your own pages matter here. So does whether they are accessible. OpenAI’s OAI-SearchBot is separate from the GPTBot training crawler. Blocking the search crawler can limit your pages’ availability for ChatGPT search; it does not mean ChatGPT stops searching the web, or that every outdated answer comes from the same crawl problem.
That distinction matters. If ChatGPT quotes an old price, “unblock the bot” is a hypothesis to check, not a diagnosis you get for free.
I would check three things:
- Can OAI-SearchBot access the relevant page?
- Does that page state the current information clearly?
- Do the cited outside sources repeat an older or different version?
The same applies to our tip-reporting feature. A clear help document gives ChatGPT a direct source. A restaurant owner’s discussion or an independent comparison gives it another route. Neither guarantees a recommendation.
Make your current information retrievable, then inspect what the answer actually cites. Do not confuse access with control.
Turn the gap into a repeatable count
Now we can measure something more useful than two attractive screenshots.
You do not need a paid tracker to begin. You need consistent questions and a record of the answers. I would use this sequence:
- Choose 8 to 10 unbranded buyer questions. Keep them close to real purchase decisions.
- Run each question three times on both assistants, starting fresh conversations or using Temporary Chat where available.
- Keep the wording and search settings consistent, and request sources where available.
- Record whether your brand appears, its list position, the alternatives named and the sources cited.
- Repeat the same set over two weeks before treating the pattern as a working diagnosis.
For our payroll company, the set might include payroll for restaurants with tip pooling and hourly staff scheduling plus payroll, alongside variants covering the same buyer’s needs. Keep your brand name out. Asking an assistant about you tests something different from asking it to build a shortlist.
A manual check that matches the buyer view also pairs naturally with a look at GA4 Traffic acquisition for chatgpt.com visits. The referral data and the prompt results answer different questions: who arrived, and how often you appeared.
For one question repeated nine times over the tracking period, 7 appearances in Gemini and 1 in ChatGPT is a gap worth investigating. It is not proof that the same split holds across the whole market. Keep the question attached to the count.
Your unit of analysis is the buyer question, repeated by assistant. An overall average can hide the exact purchase question you keep losing.
Use a tracker to preserve the method, not replace it
Paid trackers add history and charts to the same basic exercise. That can save work. It cannot rescue a prompt set that has little to do with your buyers.
What a deck calls “coverage” may just be a large collection of questions nobody asks before buying from you. I would rather have a smaller, repeatable set connected to actual decisions than a dashboard full of flattering mentions.
The 30-prompt repeat-run method gives that work a structure: consistent buyer wording, repeated runs and comparison by assistant. With groas, I want the visibility data connected to the next piece of work, not left in a reporting tab.
For each persistent miss, write a short diagnosis:
- Which buyer question are you losing?
- Which assistant shows the gap?
- Which sources does it cite instead?
- What relevant information or mention is missing?
“Improve AI visibility” is not a brief. “Our restaurant tip-pooling page is absent from the sources used for this question” is at least something you can investigate.
Choose a workflow that turns a repeated miss into a specific source-level task. More charts are optional.
Fix the assistant tied to your buyers first
The larger gap is not automatically the more valuable one. If our payroll company appears less often in Gemini but its buyers build their shortlists in ChatGPT, Gemini may not deserve the first week of work.
I would check three things before choosing:
- Where relevant demo requests appear in GA4 and the CRM.
- Which assistant buyers actually use to research vendors.
- Which purchase questions show the clearest repeated gap.
A phone on someone’s desk is not a substitute for that evidence. Neither is my own assistant preference. The assistants use different search and retrieval paths; the priority follows the buyer, not the operator’s favourite interface.
Then weigh the work required. A Gemini gap may point toward Google-visible roundups, reviews or pages that do not associate you with restaurant payroll. A ChatGPT gap may expose a crawl block, stale pricing or a missing page answering a narrower question. Those are hypotheses to test against the citations, not universal rules about the two products.
Some fixes are mechanical. Others require earning a mention on a page you do not control. Do not put them on the same timetable just because both sit under “AI search” in the project plan.
Close the cheap gap that touches real pipeline first, then fund the slower one. My practical starting point is the assistant tied to your last ten closed deals: choose the relevant gap and work on it for 30 days.
Fix shared sources before writing two versions of reality
The retrieval paths differ, but they overlap. Both assistants can use outside discussions, review sites and roundups. Both also need accurate information about what you sell.
That means some work can address both gaps. For our payroll company, I would start with three moves:
- Make the current pricing and tip-pooling pages accessible to Googlebot and OAI-SearchBot.
- Keep the product description and numbers consistent across your site, G2 and Capterra. Give the help document enough detail to answer the buyer’s question.
- Pursue a relevant roundup that describes the feature in the language buyers use.
The aim is not to paste an identical sales paragraph everywhere and call it authority. It is to stop different sources from describing different products, prices or capabilities.
I would connect that work to tracked prompts for brand mentions across ChatGPT, Claude and Perplexity, then use the 3-pass check for whether ChatGPT recommends your business to keep the diagnosis repeatable.
Fix the source both assistants can read, not the wording of one answer. You cannot edit the answer into staying fixed. You can improve the information available for the next one.
Know when the gap is not worth a rewrite
This is not a reason for every business to stop what it is doing and rebuild its website around chat answers.
If your buyers do not use assistants to compare vendors, the immediate business case is weak. If your tracked result is 4 of 9 versus 5 of 9, you do not yet have a compelling split to explain. Keep counting before assigning a project.
Referral data helps prioritize, but I would not make a zero in GA4 the whole decision. The question remains whether this research behaviour matters to your buyers. Use the traffic data alongside the CRM and those buyer conversations, not instead of them.
For a gap worth pursuing, keep the work narrow: pick five buyer prompts, address one relevant source per week, and rerun the same set for four weeks before judging the effort. Record what changed so you are not comparing a new prompt set with an old one and calling the difference progress.
Do not rewrite on the strength of a small split or one missing mention. Spend the hours where there is a buyer, a repeated gap and a plausible fix.
Test the model on a different business
Now take a commercial landscaping company chasing property managers. The buyer question is commercial landscaper for multi-site retail in Phoenix with irrigation repair.
I would start exactly where we started with payroll: repeat the question on both assistants and record the names and citations. Suppose the company appears in 6 of 9 ChatGPT runs and 1 of 9 Gemini runs.
That tells me where the observed gap sits. It does not yet tell me why, or whether to fix it first.
Next I would inspect the sources. Perhaps Gemini repeatedly cites local roundups and Google-visible service pages that name competitors. Perhaps ChatGPT retrieves the company’s service-area page and an outside discussion that names it. The source paths differ first; the shortlists differ second.
Then I would check the buyer. If the property managers behind recent deals use Gemini for this research, the Gemini gap moves to the front. The work follows the evidence: improve the relevant service page, examine the review presence and pursue a roundup that describes the actual service. If the split runs the other way, inspect ChatGPT’s cited sources and crawl access before assuming a technical block.
Either way, fix shared inconsistencies next: the same service area, service description and numbers wherever the business is described. The model travels because it starts with sources and buyers, not a payroll-specific trick.
I still like a clean Google Ads build. But buyers can shortlist before they click. Count frequency per assistant for two weeks, fix the gap tied to your last ten deals, then improve the sources both can read. That is the order, whether you sell payroll, landscaping or anything buyers compare in a chat window.

