01 Method
What I looked at
On 15 August 2026, I ran two queries — "best business bank account for startups" and "Brex vs Ramp" — across five surfaces: Google's organic top 10, Google's AI Overview, ChatGPT, Perplexity, and Gemini. Same wording on every surface, fresh sessions, Google results targeted to the US (&gl=us&hl=en). For every AI surface, I recorded which sources were cited.
If you run marketing at a company that shows up in comparisons like these, this matters because the AI engines are already assembling their own answer, and they are not drawing on the same material to do it. Ranking well on Google does not mean an AI engine will cite you. Not ranking does not mean it won't.
A few limits, stated plainly. This is two queries, run once, on one date — a snapshot, not a longitudinal record. AI answers are non-deterministic; running the same query again may return different sources. The Google results are US-targeted but were viewed from Spain. Google's AI Overview panel says it draws on "10 sites" but only shows a partial list — six were visible for Q1, three for Q2. The rest are hidden and aren't reflected here.
02 The data
The sources, by engine
Q1 — "best business bank account for startups"
| Surface | Sources cited |
|---|---|
| Google organic (9 results, ads excluded) | nerdwallet.com, bankofamerica.com, brex.com, reddit.com, mercury.com, nav.com, rho.co, smartasset.com, youtube.com. (Revolut appeared only as a paid ad — no organic listing.) |
| Google AI Overview | nerdwallet.com (4 separate claims), investopedia.com, nav.com, bankofamerica.com, rho.co, brex.com. Panel claims "10 sites"; 6 were visible. |
| ChatGPT | mercury.com — cited twice via a "Mercury +1" chip (one bundled source, unnamed). Answer text also links out to Brex, Ramp, and Relay; Chase is named but not linked. |
| Perplexity | nerdwallet.com, mercury.com, wise.com (/gb — its UK blog), anna.money, found.com, getholdings.com, fitsmallbusiness.com, clearvaluelending.com, superframeworks.com, wildfront.co |
| Gemini | relayfi.com — the sole source for the entire answer. |
Q2 — "Brex vs Ramp"
| Surface | Sources cited |
|---|---|
| Google organic (9 results) | reddit.com, brex.com/versus/ramp (#2), ramp.com/versus/brex (#3), getkleercard.com, rippling.com, rho.co, fintechlabs.com, aspireapp.com, nerdwallet.com |
| Google AI Overview | getkleercard.com, nerdwallet.com, relayfi.com. Neither Brex's nor Ramp's comparison page is cited. |
| ChatGPT | nerdwallet.com, brex.com, relayfi.com. Ramp's comparison page is not cited. |
| Perplexity | brex.com/versus/ramp, ramp.com/versus/brex, ramp.com/blog/brex-vs-rho-vs-ramp, nerdwallet.com, relayfi.com, cledara.com, kruzeconsulting.com, techrepublic.com, g2.com, checkthat.ai |
| Gemini | aspireapp.com, rho.co, techrepublic.com, layer3labs.com, johngaltfinance.com. The last two rank nowhere in Google's organic results. |
03 Findings
Why the engines diverge
Relay is the throughline: cited four times across the four AI engines in this data set — more than Brex, Ramp, or NerdWallet — and every one of those citations traces back to comparison content Relay published about its competitors, not a page about Relay itself. What this doesn't establish is why. I don't have evidence Relay built that content with AI citation in mind, or that the pattern holds beyond these two queries. Two queries are enough to show the shape of something. They're not enough to call it a rule.
Ranking in Google organic doesn't carry over to AI citation. Brex built a comparison page about Ramp, and Ramp built one about Brex, and both of those pages rank in the top three organic results for "Brex vs Ramp." That ranking doesn't translate evenly into citations, though: Google's AI Overview cites neither page, and neither does Gemini. ChatGPT cites Brex's page but not Ramp's. Perplexity is the only engine of the four that cites both. Same two vendor-built pages, same query, four different outcomes. The same self-citation pattern shows up on the other query, too: in Perplexity's Q1 answer, Wise and ANNA both appear, both cited from their own blogs. The page Perplexity cited for Wise was its UK blog, which is why that answer framed Wise as a UK option rather than a US one, in response to a US-targeted query.
Gemini's Q1 answer cites exactly one source: Relay's roundup. Gemini recommended five accounts — Mercury, Relay, Rho, Chase, and Airwallex — but only three appear in that source; Rho and Airwallex don't. Brex, Bluevine, Ramp, and Wise do, and Gemini named none of them. A single citation doesn't mean the answer derives from the cited page — where the other two recommendations came from isn't visible to anyone reading the answer. One query, one date: this doesn't establish how Gemini behaves in general.
04 Recommendations
What to do about it
If you run marketing at a company that gets compared in searches like these:
Publish comparison content about your top two or three competitors. Every citation Relay picked up in this data set came from something it wrote about a competitor, not from a page about Relay itself. If your product gets evaluated head-to-head in search — bank versus bank, tool versus tool — that's the page worth writing before a competitor writes it about you.
A Google ranking doesn't guarantee an AI citation, as the Brex-Ramp comparison pages proved.
None of this is fixed. AI answers change between sessions, and one competitor publishing a single new comparison page was enough to become Gemini's entire source base for a query — there's no reason to assume that won't happen again.
This is the kind of source-by-source tracing I do for clients.
Not affiliated with any company named. Based on publicly available data as of 15 August 2026.