We Measured Who AI Actually Cites for Montreal Agency Queries

Between 27 and 28 August 2026 we ran the same set of buyer queries through ChatGPT, Perplexity and Google AI Overviews, in English and in French, and recorded not just whether our own agency was named but exactly which sources each engine used instead. Fifteen runs completed. We were quoted in three of them, and in none of the twelve queries where the person asking did not already know our name. This is the full dataset, including the parts that do not flatter us, because a study that only reports its author winning is an advertisement.

What we measured

Ten queries, chosen because they are what a real buyer types rather than what an agency wishes they typed: two brand queries, six commercial category queries, one pricing query, and two in French. Each was run against three answer engines and scored on three things — whether an AI answer appeared at all, whether we were named in it, and the complete list of sources the engine retrieved.

That third column is the one that matters. Knowing you were not cited tells you nothing actionable. Knowing that the engine instead pulled six pages from a competitor’s own website, or four directory listings, tells you exactly what it trusts in your category.

• Queries were run in a Montreal-located browser session, logged out for ChatGPT and Perplexity where possible.
• Google AI Overviews suppress under browser automation, so every Google result was re-checked in a normal personalised session before being recorded.
• Perplexity allows exactly one query anonymously before a sign-up wall; the remaining nine were run signed in.
• French queries were scored as an independent set, not as translations of the English ones.

The headline result

Both queries that quoted us carried our name in the question. On the twelve cold commercial queries — the ones a prospective client actually types when they do not yet know who to call — we appeared zero times.

The three engines do not agree with each other

This was the finding we did not expect, and it is the most useful one. The three engines were not drawing from a shared pool of trusted sources. They were running visibly different selection mechanisms, which means a single tactic will not move all three.

Google AI Overviews select on review mass

On every local commercial query, the AI Overview resolved to Google Business Profile entities, and the entities it named shared one property. We recorded the rating and review count of each business cited on "best SEO agency in Montreal for small business":

Fifteen is the floor we observed — the smallest review count on any business the engine was willing to name. Note that rating did not order the list: a 5.0 with 49 reviews ranked below a 4.8 with 214. Volume appears to matter more than average score, which is the opposite of what most businesses optimise for.

There is a hard implication here for any small agency. No amount of schema markup, llms.txt tuning or answer-block restructuring will get you into these particular answers. The gate is off-site, it is slow, and it is the single highest-leverage thing an unknown local business can work on.

Perplexity selects on something you can actually control

Perplexity behaved completely differently, and this is where the good news is. Its source sets were roughly half third-party roundups and half agencies’ own service pages — and the agencies cited off their own sites were not the market leaders. Chivalae, RightNode Media, Lenoretech, GMN Studio, Média Conceptions, Netleaf, seomontreal.io, DVCOM and Solutions M were all quoted straight from pages they published themselves, with no review mass behind them.

That means this surface is winnable with on-site work by a business that has not yet earned the reviews to compete on Google. It is the cheapest citation available to a small agency, and almost nobody is deliberately going after it.

The pattern: pages titled like the question

Reading across every page Perplexity retrieved, the cited pages had one thing in common. They were titled the way the query was typed, not the way an agency org chart is drawn:

• "Agence SEO Montréal pour PME — Google et IA" (DVCOM)
• "GEO Agency in Montreal — Certified AI SEO Experts" (Digitad)
• "Digital Marketing Agency Montreal" (RightNode Media)
• "Web Design & Branding for Small Businesses" (MTLweb)

Our own two hits fit the same rule and prove it from the other side. The only two pages of ours that any engine ever retrieved were our pricing page and our HubSpot comparison page — the only two pages we own that are shaped like a question rather than named after a service. A page called "Services" answers nothing anyone typed.

Where we came last

The sharpest loss was the query closest to our own positioning. On "generative engine optimization GEO agency Canada", Perplexity cited five Montreal competitors — Digitad, Netleaf, Vortex Solution, BlackCat SEO and Canibuy — each from a dedicated page about being a GEO agency. Our equivalent offer existed only as a section inside a broader SEO service page. Five pages that named the thing were cited; the one that mentioned it in passing was not.

Our brand query exposed a second problem we had not known about. Of the ten sources Perplexity retrieved when asked what FutureSource is, only two were us. The other eight belonged to two entirely different companies that share the name: a UK market-research consultancy, and — new to us — a futures-and-options trading platform. An engine answering a question about you while reading mostly about somebody else will hedge, and ChatGPT did exactly that, appending a caveat that our figures were self-reported and unaudited.

What we changed

Measurement is only worth the cost if it changes something. Within two days of the run we corrected a founding date that contradicted our own LinkedIn profile, rewrote a disambiguation statement that described the other company inaccurately, removed a statistic attributed to a source that did not exist, added sourced citations to 44 pages, and built the dedicated GEO page whose absence the data had made obvious.

We will re-run the identical ten queries against the identical rubric in roughly three weeks. Holding the queries constant is the only thing that turns a snapshot into a trend, and it is the difference between research and an anecdote.

What this study does not tell you

• Ten queries in one metropolitan market is a small sample. Treat the mechanisms as directional and the exact counts as illustrative.
• Answer engines are non-deterministic and personalise heavily. The same query can return a different source set an hour later.
• The fifteen-review floor is the lowest figure we observed, not a threshold Google publishes or acknowledges.
• We measured citation, not traffic. Being named in an answer and being clicked are different outcomes, and we did not attempt to measure the second.
• The study was run by an agency that competes in the market it measured. We have published the source lists so the mechanisms can be checked independently.