The Invisible FunnelHow AI agents evaluate real brands · by Accord
Powered by Accord

Research archive

Everything that is not the current release: earlier studies run under the same harness with different questions and no memory arm, the brands the agents met as competitors along the way, and two studies of offers we operate ourselves.

What the agents actually encountered

Across the 207 recorded journeys in the research archive.

67/207journeys encountered a site that refused the research agent
17/196journeys with a brand under study ended without a readable page from that brand
136decision-shaping claims came from outside a readable brand-owned page

This measures this research harness on these missions, not every personal-agent product. A journey is one synthetic customer, one task, one agent session, on the live web; a claim is one fact the agent recorded about the brand under study.

Earlier studies

BrandConsideredNever consideredSelectedNot selectedNo readable brand-owned pageBrand site refusedMission
Google Fi WirelessTelecom · earlier study45/5779%12/5721%12/5721%33/5758%0/570%0/570%Named-offer checkThree-line household plan, all-in price and phone support
Westpac logoWestpacFinancial services · earlier study18/18100%0/180%18/18100%0/180%0/180%0/180%Named-offer checkWestpac Life savings rate for a 32-year-old saver
B&You (Bouygues Telecom)Telecom · earlier study1/333%2/367%0/30%1/333%0/30%0/30%Named-offer checkForfait mobile pour un foyer de trois lignes
Capital One 360 Checking logoCapital One 360 CheckingFinancial services · earlier study1/1100%0/10%1/1100%0/10%0/10%0/10%Named-offer checkNo-fee checking account with human support

Counted the same way as the release; not the same missions, so not comparable row to row. Each count is over the journeys in which that brand was under study; the rate beside it is that count divided by those journeys, nothing more. Considered: the brand appeared among the options the agent weighed. Not selected: considered and not selected, which includes options left unresolved and options passed over for another; the breakdown by recorded outcome is on each brand page. No readable brand-owned page: the agent obtained no readable content from the brand’s own site before deciding; it may still have requested pages there and been refused. Brand site refused: the brand’s own site answered the agent with a CAPTCHA, an access-denied page or a bot wall. Mission: Named-offer check means the agent was asked whether a named offer meets typed requirements, so a selection means the offer qualified, not that it beat the market; Open-market search means the agent chose among whatever it found. The brands were given different missions; the rows are not a ranking.

Brands encountered as competitors

BrandMet inNamed as an optionSite requestedRefused by the siteSelectedRejectedUnresolved
Visible (Verizon)Google Fi Wireless study45/5188%33/5165%0/510%0/510%27/5153%18/5135%
Mint MobileGoogle Fi Wireless study44/4892%29/4860%29/4860%0/480%10/4821%29/4860%
American AirlinesTravel after telecom study10/1191%10/1191%10/1191%0/110%3/1127%5/1145%
United AirlinesTravel after telecom study10/1191%11/11100%0/110%0/110%3/1127%6/1155%
Southwest Airlines logoSouthwest AirlinesTravel after telecom study10/1191%11/11100%0/110%3/1127%3/1127%0/110%
US MobileGoogle Fi Wireless study9/1090%6/1060%0/100%1/1010%0/100%8/1080%
Target logoTargetAmazon, Temu study6/1060%10/10100%0/100%0/100%2/1020%4/1040%
Walmart logoWalmartAmazon, Temu study7/1070%10/10100%6/1060%0/100%0/100%7/1070%
Best BuyAmazon study5/683%5/683%0/60%0/60%5/683%0/60%
DysonAmazon study5/683%5/683%5/683%0/60%1/617%3/650%
Booking.comAirbnb study4/580%5/5100%0/50%0/50%0/50%3/560%
VrboAirbnb study4/580%5/5100%0/50%0/50%0/50%2/540%
Delta Air LinesTravel after telecom study4/4100%4/4100%0/40%2/450%1/425%0/40%
Amazon logoAmazonTemu study2/450%4/4100%0/40%0/40%0/40%2/450%
Aer LingusRyanair study2/450%4/4100%0/40%0/40%0/40%2/450%
VuelingRyanair study1/333%3/3100%3/3100%0/30%0/30%1/333%

Each count is over the journeys in which the agent named the brand as an option or requested its site, inside another brand’s study; the requirements were written for that study, not for this brand. Named: appeared among the options the agent weighed. Site requested: the agent asked the brand’s own site for at least one page; subtract the refused count for the journeys with a readable page. Refused: the brand’s site answered with a CAPTCHA, an access-denied page or a bot wall. Rejected and unresolved are the agent’s recorded status for the option.

Our own offers

Accord (Agentdoor) (70 journeys) · LaunchBlock (20 journeys). Kept out of every consumer-brand table.

Archive stories

Google Fi Wireless

The agent read Google Fi’s pages in all 57 journeys. In 31 it still could not decide.

In 57 recorded research runs, the agent read fi.google.com every time and support.google.com in 38. It selected Google Fi in 12, a competitor in 2, ran out of page reads in 12, and ended 31 without a recommendation because a required condition could not be confirmed from a readable page.

57 runs · named-offer check · one research-agent configuration · no purchase made

Mint Mobile

Mint Mobile refused the agent in 29 of 29 visits. It was never selected.

In the 57-journey Google Fi study, the agent named Mint Mobile as an option in 44 runs and requested mintmobile.com in 29. All 29 requests were refused.

44 runs · met inside a named-offer check of Google Fi · one research-agent configuration · no purchase made

Westpac logoWestpac

Eighteen journeys, eighteen selections, two to six page reads. Westpac, from its own pages.

In 18 recorded research runs on 18 September, the agent established the total achievable rate on Westpac Life from westpac.com.au alone, in two to six page reads, and selected it every time. No third-party page entered any decision.

18 runs · named-offer check · one research-agent configuration · no account opened