From the Gravity Funnel to Fly By Gravity
The Gravity Funnel named the mechanism: pull, not push. But a pull has a source. Naming it meant turning away from the act of capturing demand and toward the thing that generates it. The attraction had to be named on its own terms. That name was Fly By Gravity.
A referral source's conversion rate measures arrival, not persuasion.
A source's conversion rate rises when the deciding moves off the surfaces that report it. That makes the AI referral ratio an index of what analytics stopped seeing, and a poor basis for deciding what to fund.
Brainlabs gave Digiday fourteen months of Google Analytics data covering 54 advertisers, and the number everyone repeated was the smallest one in the set. Organic sessions across the sample fell 10.5%, from 140.1 million to 125.4 million. Sessions from AI platforms rose 163%, key events from those platforms rose 335%, and referrals from ChatGPT, Copilot, Gemini and Perplexity produced key events at 1.5 times the rate of organic search traffic. The reading that followed was that AI is the better channel. The data will not carry that reading. A conversion rate measured by source records when a visitor showed up, and what these figures describe is a buying journey whose middle has moved somewhere the instrument cannot see.
What does a 1.5x conversion rate on AI referrals actually measure?
A conversion rate measured by referral source records how late in the buying journey a visitor arrived, not how well that source persuaded them. Google's own analytics documentation is explicit about the limit: attribution is the act of assigning credit for user actions to the ads, clicks and factors along the path to completing the action. Credit can only be assigned to interactions the system observed. Everything that happened off the measured surface is not scored badly, it is not scored at all.
Call the resulting artefact arrival credit: the share of persuasion a measurement system hands to the last visible source because it holds no record of the earlier ones. Arrival credit has one property that decides everything downstream. It rises as observability falls. A person who spent forty minutes comparing options and asking follow-up questions inside a chat interface, then clicked one link and bought, registers as a spectacularly efficient referral. Nothing in that record describes persuasion. It describes lateness, and the record is flattering in exact proportion to how much of the work it failed to capture.
The same arithmetic has always run in reverse, which is why this is a known error rather than a new one. A source that reaches people early, while they are still forming the question, converts poorly on the visit and well over the relationship. Its conversion rate is a readout of the population it attracts, not a measure of its persuasive force. Branded search is the standing example: it converts far better than unbranded search, and no serious analyst believes the search box built the brand. The conversion rate of a source and the contribution of a source are different quantities, and only one of them is easy to put in a report.
How large is the AI referral base next to the organic decline?
The AI referral base in the Brainlabs sample is a small fraction of the organic volume that disappeared, and the published account of that volume does not reconcile with itself. Digiday describes the sample's organic sessions falling from around 20 million a month to less than 15 million, and describes the same decline as 140.1 million to 125.4 million across fourteen months. Those two statements are not the same quantity: 125.4 million over fourteen months is about 9 million a month. Nothing in the reporting explains the gap, and an argument about people misreading numbers has to say so rather than quietly pick the convenient figure.
The conclusion survives either basis, which is the useful thing about doing the division. AI sessions in the sample ran at roughly 200,000 a month. Set against 15 million that is 1.3% of organic sessions; set against 9 million it is 2.2%. Apply the 1.5x conversion advantage and referrals from AI platforms are producing somewhere between two and three and a half percent as many key events as organic search. That is the quantity the celebrated ratio is attached to. A 163% rise on the smaller number and a 10.5% fall on the larger one are not two halves of one channel shift, and percentage growth on a small base is the oldest way to make a rounding error look like a destination.
The evidence is also thinner than the confidence being placed on it. Brainlabs published an analysis of its own client Google Analytics data through a trade publication, not a study with a stated methodology, a control group or a published sample frame. It covers 24 US firms, 24 British brands, six clients elsewhere and 19 sectors, weighted toward large advertisers and retail. Its finding that 46 of 54 clients lost organic sessions is broad, consistent and worth taking seriously as a description of decline. It is not a measurement of where persuasion went, and it does not claim to be.
If the persuasion did not happen on the site, where did it happen?
Google's search documentation places part of the answer inside the organic number itself. Sites appearing in AI features such as AI Overviews and AI Mode are included in the overall search traffic in Search Console, reported within the Web search type. An answer engine that sends a click therefore reports as organic search, and the same engine reports as nothing at all when it answers without sending one. Both the 10.5% organic decline and the organic conversion rate it is being compared against are already partly measurements of AI usage, mixed in by an instrument with no facility for separating them.
Pew Research Center supplies the behavioural half, with a boundary worth naming. Pew analysed the browsing data of US adults who agreed to share it and found that users who encountered an AI summary clicked a traditional search result in 8% of all visits, against 15% for users who did not, with clicks on links inside the summary occurring in just 1% of visits. Pew measured a search results page carrying a summary, not a conversation inside an assistant. It supports one claim and not the larger one: reading and comparing can now complete without producing a visit. What happens inside a chat interface is not covered by it, and no public dataset covers it.
What remains in the analytics is the last observable footstep of a walk that was mostly invisible. A referral from an AI platform is a person arriving after the comparison is finished. An organic session is now a mixture of people arriving after a summary answered part of their question and people arriving cold. Neither figure carries information about which material did the convincing, and arrival credit is the name for the difference between what the instrument reports and what actually took place.
Is the 1.5x figure simply proof that AI visitors have higher intent, and should budget follow it?
The strongest case for moving budget toward AI referrals is that higher intent is a real property with real commercial value, and buyers have always paid more for it. Media buying has priced intent for decades. If referrals from ChatGPT, Copilot, Gemini and Perplexity genuinely arrive later in the buying journey, they are worth more per visit, and a rational buyer moves marginal money toward whatever produces them. A business that ignores visibility in those systems because today's volume is small is making a straightforward forecasting mistake, and the 335% rise in key events from AI platforms is a fair warning about direction. That case is correct about the value of the arrival.
The marginal argument requires something to buy, and in this instance there is nothing. Intent-bearing inventory has always been purchasable on its own: an advertiser can bid on the late query without touching whatever created the demand behind it. A citation inside an assistant has no auction, no placement and no inventory. The only input that produces one is published material a system can read, which is the identical input that produced the organic session it is being compared against. There is no margin to move along, because the two budget lines are not two channels competing for one buyer. They are one production process measured at two exits.
What survives the objection is a change in the form of the work rather than a transfer of funding away from it: material specific enough to be quoted, current enough to be trusted, structured so an answer can be lifted from it whole. The transfer is the failure mode, and it runs through the single lever available. Cut spend on published material and the first number to fall is organic sessions, which is already falling for reasons everyone has agreed to blame on AI. The AI referral ratio holds or improves, because the people who still arrive that way still arrive late. The instrument reports the cut as an efficiency gain. The delay between the decision and the visible consequence is the time it takes indexed material to age out of what these systems read, nobody has measured that interval, and it is long enough that the loop does not close.
What measurement would settle this?
The claim that AI referrals collect arrival credit rather than earn conversions is falsifiable, and testing it needs no new instrumentation. Choose a set of topics, stop refreshing and extending the published material behind them, change nothing else, and watch two numbers together: total key events across the business, and the conversion rate of AI referrals. The prediction is specific. The ratio holds or rises while the totals fall. If the ratio falls alongside the volume, then those referrals were being produced by something other than the material, arrival credit is not what is inflating them, and this argument is wrong.
A ratio cannot be read against itself, which is why the pairing is the whole test. Arrival credit moves a source's conversion rate and a business's total conversions in opposite directions, so the signature of the error is a metric improving while the thing it is supposed to represent shrinks. Any business that has ever watched branded search efficiency improve during a demand decline has already seen this signature and had no name for it.
The honest result of this test in most businesses will be unknown, and unknown is a legitimate reading that should be recorded as unknown rather than rendered as zero. Sample sizes are small, seasonality is real, and a ratio calculated on roughly 200,000 monthly sessions and reported to one decimal place deserves less confidence than it is being given. The expensive mistake available here is not a wrong number. It is a confident number attached to the wrong quantity.
Why will an instrument that measures arrival defund the work that persuades?
An instrument that records arrivals will defund alignment and report the defunding as efficiency, because alignment produces no event and arrival does. Alignment between a message and a customer already in motion is temporary. It holds while trust, relevance and timing hold, and it decays without the message ever becoming bad. Nothing fires when it decays. Something fires every time somebody shows up, and that asymmetry is not a flaw in one analytics package. It is what arrival-based measurement is.
The funnel model's accounting made this error a generation earlier by treating whichever touch preceded the purchase as the one that did the advancing, for no better reason than that it was the touch that left a record. The Marketing Helix describes the same customer differently: already in motion, already deciding, pulling in the messages that hold trust, relevance and timing. The measurement consequence is the part that matters here. Material being read and summarised inside systems that send no click is still holding alignment and still generating no arrival, so the more of the work that moves into those systems, the more of it the instrument writes down as costing money and producing nothing.
The AI referral figure is that error in newer clothes, and it will be harder to resist this time because the number is larger and the channel is new. A 1.5x conversion rate does not identify a better channel. It identifies a later one.
A conversion rate by source is a census of who arrives, not a verdict on who persuaded.
The 1.5x ratio has one property that should end its career as a budget input: it improves as observability falls. Every further shift of the deciding into interfaces that publish no record will push it higher, and every increase will read on a dashboard as a channel performing better. A number that goes up when measurement gets worse is not a performance metric. It is an index of how much of the work is now happening where nobody is counting it.
Further reading
- Digiday, reporting Brainlabs' Google Analytics analysis of 54 advertisers, the study this essay is grounded in digiday.com/marketing/in-graphic-detail-how-ai-search-has-impacted-the-web-traffic-of-over-50-advertisers
- Pew Research Center browsing-data study, evidence that clicks fall by roughly half when an AI summary appears pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links
- Google Analytics Help, the primary definition of attribution used in the argument about what credit can and cannot be assigned to support.google.com/analytics/answer/10596866
- Marketing Helix, the structural comparison this essay's measurement argument extends marketinghelix.com/helix-vs-funnel
Every source above was fetched and a verbatim phrase confirmed on the page before this essay published. Nothing here is paraphrased from memory.