GuideOct 11, 202611 min read

How to See Which Brands ChatGPT, Gemini and Perplexity Recommend Instead of You (and Fix It)

A manual check in one engine shows you one model on one day. Here is how to track who the engines name instead of you, read the result by buying stage, and fix the gaps in an order that holds up.

• ~4xspread in visibility for one brand across four engines, same prompts, same week
• 1,000distinct domains cited across 599 answers in a single weekly report
• 94of roughly 150 buyer questions where one software review directory was cited

To see which brands AI engines recommend instead of you, run one fixed set of buyer questions through ChatGPT, Claude, Gemini and Perplexity on a schedule, record who is named, cited and recommended in each answer, and trace the sources behind each recommendation. A manual check shows one engine on one day; a tracked set shows the gap by stage and where it comes from.

In our October 2026 scan I put the exact question most people arrive here with, "How do I even know if AI assistants are recommending my competitors instead of us?", to ChatGPT, Claude, Gemini and Perplexity, and all four came back with the same first move: write down 20 to 50 questions your buyers would ask, run them in fresh sessions across the engines, log who gets named and which sources are linked underneath, and repeat it every week or two. Perplexity's version added the caveat the other three left out, that a small sample can be inconsistent, and ChatGPT's went a step further and separated "mentioned" from "recommended", which is the distinction most people miss on their first pass.

That's a fair method and it's the one I'd start with, at least for the first afternoon. The trouble is what happens after it. Done by hand in one browser tab, the method produces a screenshot, and in our own data the same brand's visibility differed by close to a factor of 4 depending on which tab you happened to have open.

So what follows is the manual method and exactly where it stops being useful, then what a multi-engine tracker measures and how to read those numbers by buying stage, then how to turn the gap into a fix list in an order that holds up, with short notes for enterprise teams, founders, agencies and govtech vendors, and finally what it takes to close the loop so the next report shows whether any of it worked.

Can you just ask ChatGPT which brands it recommends?

Yes, and it's the right first step. Open a fresh session in each engine, ask 10 to 20 of the questions a buyer in your category would actually type, and write down which brands appear, whether each one is a recommendation or a passing mention, and which URLs are linked under the answer. What you get is one sample from one model on one day, which is enough to see that the problem exists and not enough to size it.

The first limit is the engine. On the same 150 prompts in the same week, our own brand scored 2.7 out of 100 on Gemini, 3.6 on ChatGPT, 8.5 on Perplexity and 10.2 on Claude, with the mention rate running from 2.7% on Gemini to 12% on Claude. That isn't unusual. Each model retrieves from a different set of pages and weights them differently, so a founder who checks Gemini concludes the brand is invisible and a founder who checks Claude concludes it's doing fine, and both are describing one engine.

The second limit is the session. Our weekly score on an unchanged prompt set ran 5.0, 4.8, 4.9, 6.2 and then 5.1 across the five reports from 10 September to 8 October 2026, with no change to the prompt set that would explain the jump in the fourth week. One competitor-tracking prompt in that set went from the brand being present in 3 of 4 engine answers to 1 of 4 inside a single week. If you'd sampled that prompt by hand on either side of the change you'd have drawn a confident conclusion from what is, as far as I can tell, variance.

The third limit is the source trail. A manual check gives you the two or three links under one answer. Across the 599 answers in one of our weekly reports, the engines cited 1,000 distinct domains, and the recommendations for our category leaned on a small number of them: one software review directory's Australian edition was cited in 94 of the roughly 150 questions, a software listing site in 88, and our own domain in 46. You cannot see that shape from a browser tab, and the shape is what tells you where the fix goes.

What does a multi-engine tracker actually measure?

A multi-engine tracker runs a fixed set of buyer prompts through ChatGPT, Claude, Gemini and Perplexity on a schedule and records, per answer, whether your brand was mentioned, whether your site was cited as a source, whether you were recommended, and where you sat in any ranked list. Rolled up across the set, those become mention rate, citation share, recommended rate and prompt coverage, each broken out by engine and by stage.

The four are worth keeping separate because they fail separately. Mention rate is the share of answers where the brand turns up in the prose at all. Citation share counts something narrower, the answers where your own domain sits among the linked sources. Recommended rate is stricter again, the answers where the model actually put you forward rather than listing you in passing, and it's the number ChatGPT itself told us to watch. Prompt coverage works the other way round, asking what share of the prompts you appear in on at least one engine, and it's the one that maps most directly onto a content plan: every uncovered prompt is a question your site doesn't answer well enough for any engine to use.

The gap between the first two is the most useful single reading a tracker gives you. In our October 2026 report the site was cited as a source in 10.2% of answers while the brand was named in only 7%, and in the Solution Research stage it was 12.1% cited against 6.5% named. The model read a page of ours, used it, and still didn't say the name. That is an attribution gap, and the fix for it is entity work, the same legal name and category description on every page and in the Organization schema, which is a different job from the coverage gap you'd fix by writing pages. I've gone through the mechanics of that in why AI models skip brands they can't place, and the score itself is unpacked in what an AI visibility score measures.

How do you read the results by funnel stage?

Read a tracker's results one buying stage at a time, and compare each stage against competitors' rates in that same stage, never against your own rate in a different stage. Presence climbs through the funnel for every brand in a category, the leader included, because later-stage questions name fewer brands and the model has less room to leave you out.

Rushana's line on calls is that the questions people are asking are the new keywords, and the stage is the part of the keyword that tells you what the question is for. We group prompts into four stages, and the reading is different for each.

Problem Recognition prompts are the ones where a buyer hasn't named a category yet, "why does ChatGPT keep recommending our competitors" being the obvious one. On our own competitor-tracking prompts at this stage we were present in 0% of answers and the most-present rival in 30%. Nobody is shortlisted here, so the number to care about is whether any brand in your category is being named at all, because if one is, that brand is becoming the default explanation.

Solution Research prompts name the category and ask what exists in it. This is where directories and "best tools" roundups carry most of the weight, and it is the stage where our citation rate ran furthest ahead of our mention rate. A gap here is usually off-site.

Vendor Evaluation prompts compare named options for a named situation, and they're the ones closest to a deal. The top rival on our set was present in 75% of these answers and we were present in 8.3%. One well-built page that answers the specific comparison moves this stage more than anything else we've tried, because so few brands get named in each answer.

Competitive Intelligence prompts ask how to beat or switch from a named rival. They're low volume and high intent, and a brand that is cited here without being recommended is being used as evidence in someone else's comparison. Read them for sentiment as much as presence.

BrandViz.AI's own 17 competitor-tracking prompts, 4 engines, Australia market, report of 8 October 2026. Presence means named or cited in the answer.

Two cautions on reading a table like that.

A stage with four prompts in it is a small sample, so a single prompt flipping moves the rate by a quarter, and the honest reading is the direction over several reports rather than the week.

A prompt where no brand and no source is named by any engine is uncapturable for now, so it belongs out of the fix list entirely rather than in it as a zero.

How do you turn a competitor-citation gap into a prioritised fix list?

Take every prompt where a competitor is present and you are absent, drop the ones where no brand and no source is named at all, sort what's left by stage with Vendor Evaluation first, and then for each prompt read the cited sources to decide which of three layers the fix sits in: the technical layer, the content on your own site, or off-site presence someone has to go and earn.

The three layers are Rushana's frame and I haven't found a gap yet that doesn't land in one of them. Technical is whether the engines can fetch and parse your pages and resolve you to one entity. Content is whether a page on your site answers the prompt directly enough to be quoted. Off-site is the directories, reviews and third-party write-ups the models lean on when they decide who to recommend, which someone in the team has to go and place. The cited sources under each uncovered prompt tell you which layer you're in, and that's the order I'd check them.

The OAI-SearchBot and robots.txt details are from OpenAI's crawler documentation, read October 2026.

Vendor Evaluation goes first for the reason the stage table showed: a prompt like "which platform handles X for a 20-person team" sits nearest the deal, and a page that answers it has little competition in the answer. Problem Recognition prompts are worth a page each too, since a brand that explains the problem tends to get named when the buyer moves on to the category, but they pay off later and they're lower in the list for that reason. The way I think about it is that the engine is your sales rep, and a sales rep needs the capability answers in the order a buyer asks them.

What changes depending on who you are?

The method is the same for everyone. What changes is which stage you read first, how many prompts you need, and who gets the fix list.

Enterprise marketing teams

You have the prompt volume and usually several markets, so the question is less "are we visible" and more "which model, which market, which stage". Read by engine first, because a blended score across four models will hide the one where a rival has pulled ahead. Entity resolution is the fix that comes up most for larger organisations, since the same company tends to be described six different ways across product lines and regional sites, and the models settle on the version a directory wrote. The fix list usually needs a technical owner from day one.

Founders

You found this by searching your own company and watching three rivals come up, and you don't have a team to hand the list to. Keep the prompt set small, 30 to 50 for one market, and read Vendor Evaluation first because that's where the deal is. The crawler-access check and the entity fix are both a morning's work and they move every prompt at once, so do those before writing a single page.

Agencies

You're running this across a client list, so the thing to standardise is the stage split and the source trail, not the prompts themselves, which differ by category. The source list is also your pitch: showing a client that one review directory carries most of the recommendations in their category is a clearer brief than a score. Rushana's comparison is that reading the global top sources for a specific client is like dressing for Brisbane by looking at the global temperature, and the same goes for reusing one client's fix list on another.

GovTech vendors

Your buyers ask long, constrained questions, "which [category] vendors meet [standard] for a state agency", and the engines answer them from a thin set of sources because very little is published. That makes prompt coverage the number to watch, and it means a single page that answers the constrained question directly, with the certifications you actually hold stated plainly and the ones you don't hold left out, tends to be cited quickly because nothing else is competing for it, at least from what we're seeing. Don't let anyone write you into a compliance claim you can't back; the models will repeat it.

How do execution platforms close the loop?

An execution platform takes the prioritised fix list and does the work: it writes the schema and entity fixes, drafts the pages that answer the uncovered prompts, places them in your CMS or codebase as drafts for someone on your team to approve, and then rescans the same prompt set on the next report so you can see which items moved which engine. The tracking layer is the same as a monitoring tool's; the difference is a build layer and a delivery layer on top of it.

The reason this matters for the method above is the last step. A fix list nobody executes stays a list, and a fix that ships without a rescan can't be told apart from the week-to-week noise I described earlier. Running the same prompts weekly after each batch is what turns that noise into a signal, because a change that holds across three reports after a specific item shipped is a result and a change that appears in one report is variance. It's a lot of small strategies and pins that together add up to visibility, and the loop is how you find out which pins held. The tracking versus execution guide sets out the three classes of platform on exactly this question, and how to evaluate an AI visibility platform has the scorecard if you're comparing options.

That loop is what we built BrandViz.AI to run. Starter is $200 AUD a month for 50 tracked queries in one market, refreshed monthly, with the Engine shipping around 8 items to your site each month; Growth is $1,000 AUD a month for 150 queries across up to 3 markets, refreshed weekly, with around 30 items a month. Nothing goes live without your approval. If you'd rather start with the manual afternoon, the free audit reads your live site the way the crawlers do and hands you the technical half of the fix list first, which is the half most teams have never checked.

Common questions about tracking competitor AI citations