We asked three AI assistants for a vendor. Here is what they searched first.

Key takeaways

• We put ten buyer prompts through ChatGPT, Perplexity and Gemini on September 16, 2026, and ran the Gemini set again on September 21. Every search the assistants ran and every site they cited is in the table below.
• The searches are the phrases on a vendor's page, almost word for word. Gemini turned "who can automate Buildium for my property management company" into "hire buildium automation expert". ChatGPT turned the after-hours prompt into "after hours property management maintenance call answering service emergency triage Canada".
• Where a category already has known vendors, the assistant searched brand names it already held, in quotation marks, before it searched the category. The shortlist existed before the search did.
• Five of the ten prompts triggered no search at all on Gemini. Those answers came from what the model already believed, and no page you publish this month changes them.
• One run is an anecdote. Between the two Gemini runs, one prompt stopped searching and another kept citing nezamai.com but dropped the firm from the answer. Measure the rate, monthly, on the same prompts.

The test

What AI assistants search before they recommend a vendor is readable, because each of the three shows its work if you know where to look. Perplexity lists its searches under "Researched" as it answers. ChatGPT writes its searches into the response stream as a `queries` array, which the browser's developer tools show under the Network tab. Gemini, called through its API with Search grounding switched on, returns the searches it ran as a field called `webSearchQueries`.

So we wrote ten prompts the way our buyers type them. Four as a Toronto property manager, three as an Ontario brokerage owner, two as a GTA plumbing company owner, and one as anyone checking the brand. We ran all ten through the three assistants on September 16, 2026, from a fresh thread each time, and ran the Gemini set again on September 21 to see what moved in five days.

Two caveats before the table. The ChatGPT run was from a signed-in account with memory on, and its first answer mentioned projects of ours by name, so treat its searches as biased toward what it already knew about us. Perplexity, on the free plan, showed no such effect. And Gemini was called through the API rather than the app, so it is the search behaviour of the model, not of the product's interface.

| Prompt | Assistant, date | What it searched |

|---|---|---|

| best AI answering service for a property management company in Toronto | Gemini, Sept 16 | "AI phone answering service property management emergency maintenance leasing", then "best AI answering service property management" |

| | Gemini, Sept 21 | nothing |

| | ChatGPT, Sept 16 | five searches, four of them `site:` searches on eliseai.com, colleen.ai, buildium.com and rentvine.com |

| who can automate Buildium for my property management company | Gemini, Sept 16 | "automate Buildium property management integrations consultants", then "Buildium automation agency consultant property management" |

| | Gemini, Sept 21 | "Buildium automation agency consultant integrations", then "hire buildium automation expert" |

| | ChatGPT, Sept 16 | "Buildium automation API Zapier property management consultants Buildium integrations 2026", then a `site:buildium.com` search |

| how do I cover after hours maintenance calls without hiring a night shift | Gemini, both dates | nothing |

| | ChatGPT, Sept 16 | "after hours property management maintenance call answering service emergency triage Canada property management" |

| AI consultant for property management companies in Toronto | Gemini, Sept 16 | "AI consultant property management Toronto", then "proptech AI consulting Toronto real estate property management" |

| | Gemini, Sept 21 | "proptech consultant Toronto property management", then "AI consulting property management Toronto" |

| | ChatGPT, Sept 16 | "Toronto AI consultant property management companies automation real estate", then "Toronto property management AI consulting automation" |

| Follow Up Boss automation consultant | Gemini, Sept 16 | "Follow Up Boss certified partners", then "Follow Up Boss automation consultant" |

| | Gemini, Sept 21 | "Follow Up Boss automation consultant agency certified partner", then "hire Follow Up Boss automation specialist" |

| | ChatGPT, Sept 16 | "Follow Up Boss certified consultant automation setup integrations consultant", then "Follow Up Boss consultant automation Zapier integrations real estate" |

| answering service for plumbers that books jobs into Jobber | Gemini, Sept 16 | four searches, two of them vendor names in quotes: "Pink Callers" and "Jill's Office" |

| | Gemini, Sept 21 | six searches, four of them vendor names in quotes: "Ruby", "Jill's Office", "Smith.ai" and "Pink Callers" |

| | ChatGPT, Sept 16 | three searches, one of them `site:jobber.com` |

| Nezam AI review | Gemini, both dates | "Nezam AI review", then "Nezam AI" |

| | ChatGPT, Sept 16 | five searches, including the founder's name and the domain variants |

nezamai.com was cited on two of those seven prompts. On the Toronto consultant prompt all three assistants cited it on September 16, and Gemini cited it again on September 21 without naming the firm in the answer. On the brand check all three cited it on every run. On the other five, nothing from our site was read.

Three of the ten prompts are left off the table because Gemini ran no search on either date and the ChatGPT searches were of the same shape as the rows above: AI lead follow-up for an Ontario brokerage, an AI receptionist for a real estate office at night, and an AI receptionist for a plumbing company in Toronto. The full capture, with every source cited per prompt, is the log we re-run on the 16th of every month as part of the search work.

The search is the page's phrase

Read the right-hand column again and you are reading page titles. "Buildium automation agency consultant." "Follow Up Boss automation consultant." "Hire buildium automation expert." Each is a service, then a buyer's word for the person who sells it. None of them is a question, and none of them is the sentence the buyer typed.

Ahrefs measured this across 1.4 million ChatGPT prompts: the median similarity between a cited page's title and the prompt was 0.602 against 0.484 for pages that were retrieved and not cited, and the match rose to 0.656 when the title was compared with the assistant's own sub-queries rather than the prompt. The page that gets cited is the page whose title reads like the search the model ran.

That is why the week after the first run we built four pages named after the searches rather than after our services: after-hours maintenance call handling for property managers, Buildium automation, Follow Up Boss automation and LeadSimple automation. Five days later none of them is cited. That is the expected result. They were indexed this week, and the pages the assistants read instead have been there for years. The honest claim is that the phrase is now on a page, and the October run will say whether that was enough.

The shortlist existed before the search

The Jobber prompt is the one to sit with. On September 21 Gemini ran six searches for it, and four of them were a vendor's name in quotation marks followed by the word Jobber. It did not search for answering services and find Ruby, Smith.ai, Jill's Office and Pink Callers. It searched for them, by name, then filled in the category around them. Which is why the checks a contractor should run come before the vendor list, and we wrote them down.

Suganthan Mohanadasan found the same thing in ChatGPT and published it in August: in 21 of 27 conversations the first search already contained brand names the user had never typed, and a brand that appeared in ChatGPT's own search was cited 68.9 percent of the time against 2.1 percent for a page that was merely fetched. He calls the numbers directional, one account's patterns. A larger study from Genezio, 220,193 ChatGPT fan-outs published in June, put it at 97 percent: of the brands ChatGPT searched for by name, 97 percent survived into the final answer.

Our ten prompts agree with both. Which means the page you publish decides whether you are cited once you are on the list, and something else decides whether you are on it. That something is the set of pages about you that you did not write. On the brand prompt, Gemini cited our Clutch profile on both runs. It is one of a handful of third-party pages about the firm, we did nothing to it for that prompt, and it was read before most of our own site was.

Half the prompts never searched

On Gemini, four of the ten prompts produced no search at all on either date: the after-hours prompt, the brokerage lead follow-up prompt, and both receptionist prompts. On September 21 a fifth joined them, the Toronto answering-service prompt that had run two searches on September 16. The answers were long, confident and full of vendor names. EliseAI led the answering-service answer on both dates, and on the second date not one page on the web was consulted to write it.

For those prompts nothing on a website matters this month, because the model is answering from what it read before its training cutoff. The way onto that list is the same as the last section, done earlier: coverage, comparisons, reviews, directory pages, written by other people, sitting on the web long enough to be in the next model's training run. It is the slowest lever there is, and on half the prompts it is the only one.

What the brand check finds

Every assistant, on every run, searched "Nezam AI review" and then "Nezam AI". This is the follow-up search, the one that happens once a model has a name and wants to know whether to trust it, and what it finds is worth looking at closely.

Perplexity decided "Nezam AI" refers to three different products: an ERP platform at a similar domain, an Arabic e-commerce builder, and us. It described the consultancy accurately, down to the two-week audit and the four industries, and then scored the group one out of five for independent customer evidence and one out of five for pricing transparency. Gemini on September 16 gave the firm a footprint in Saudi Arabia and a practice in higher education and the public sector, none of which exist. Five days later it described us correctly. Nothing changed on our site in between.

That is what a search on a shared name returns when almost every page about a company is the company's own. We have done the two things a site can do about it: a proof page that lists every figure we publish with its window, its source and what we do not claim, and a process page that says what an engagement actually runs like and what the client owns. The rest is reviews, from clients, in their words, opening with the words a buyer searches. We will not manufacture them, and the reason is on this page: the assistants already say "mostly self-reported results", and a fake review is the fastest way to earn that line permanently.

The assistant is a Google user

One more row, from a different log. Our Search Console export for the 90 days to September 16 contains this query, verbatim, at an average position of 10.2 with five impressions and no clicks:

how to automate rent collection for landlords you are answering a software buyer's question using live web search. search the web, then recommend specific named products and say why, citing what the sources say. do not ask clarifying questions.

That is an agent's instructions, pasted into Google by whatever tool was running it. Somebody built a software-buying assistant on top of a plain Google search, and the search landed on our site at position ten. It is the plainest evidence we have that there is no separate machine to optimise for. Under at least some of these tools, the retrieval is a web search, the page that ranks for the phrase is the page that gets read, and a page-one ranking with no click is exactly what being read by an agent looks like in your analytics.

Run it on your own prompts

Write ten prompts the way your buyer types them, including one that is just your company name plus the word review. Run each through Perplexity and read the "Researched" list. Run each through ChatGPT with the developer tools open on the Network tab, find the conversation request, and search its response for `queries`. If you can call Gemini's API, switch on the `google_search` tool and read `webSearchQueries` from the grounding metadata. Log the searches, log the sites cited, and run the set three times per assistant before believing any of it, because our two Gemini runs five days apart disagreed on two prompts out of ten.

The list of searches you get back is the list of pages to have. Ours is above, and the October run goes next to it on the 16th.