Self-reported attribution asks customers how they discovered or evaluated your brand, including whether an AI assistant played a role. Pair those answers with referral and CRM data. It can reveal influences absent from click logs, but recall errors and overlapping touchpoints mean it is not a complete attribution record.
Self-reported attribution is a survey question on your highest-intent forms that asks how a buyer found you, with AI as a named option. It exists because AI referrals lose their referrer data and get filed as direct or branded search, so your analytics structurally undercount them. The data is directional, not exact. Built well, it is the single best signal you have for a revenue channel cookies cannot see.
This is not a workaround you run instead of analytics. It is a second instrument that reads what the first one cannot. Below is how to build it, when to ask, how to keep the answers honest, and how to reconcile what people tell you against what your visibility data already shows.
- AI referrals arrive without referrer headers and get misclassified as direct or laundered into branded search, so GA4 undercounts them by design.
- A self-reported attribution question on high-intent forms captures discovery your analytics misses, at completion rates of 45% to 85% when asked post-conversion.
- Structured channel options beat a single free-text box. Add a short "what did you ask" follow-up to confirm AI discovery rather than guess at it.
- The data is biased toward recent and well-known channels. Randomize options, reconcile against your model-level visibility, and graduate to a lift test once you run several paid channels.
Why your analytics will always miss part of AI's influence
Two mechanics make AI a measurement blind spot, and neither is fixable with a tag. The first is referrer stripping. When someone taps a link inside the ChatGPT app or an AI browser, the link often opens in an in-app browser that drops the referrer before the request leaves the device. Wheelhouse DMG's server-log comparison found one case where Gemini on iOS sent 56 real visits and GA4 recorded 5 of them. The second is branded-search laundering: an AI model recommends you by name, the buyer Googles that name, and the session lands as organic search. The AI's role disappears entirely.
Parse tracks AI visibility across ChatGPT and Google AI Mode, covering 3.1 million indexed prompt responses across 550K+ brands. That upstream view is what tells you the influence exists. The survey is what tells you it reached a buyer. You need both because the click in between is the part that vanishes.
The numbers above: 89% of brands cannot properly attribute AI referral traffic, roughly 70% of AI sessions arrive with no referrer and get filed as direct, and even after Google added a native AI Assistant channel to GA4 in May 2026, an estimated 20% to 40% of AI-originated visits still miss it. That residual is permanent. It is why the survey is not optional.
What self-reported attribution actually measures
It measures recalled discovery, not a verified touchpoint. That distinction matters because it sets the right expectation with leadership. Rand Fishkin's argument at SparkToro is that marketers should stop chasing single-touch attribution and start tracking aggregate measurement: directional indicators of influence, watched over time, validated with lift tests. A self-reported question fits that frame. It will not tell you that this specific $40,000 deal came from Perplexity. It will tell you that AI's share of "how did you find us" moved from 3% to 11% over two quarters, and that trend is real even when any single answer is fuzzy.
So set the bar honestly. The survey is a leading indicator and a cross-check, in the same family as branded search volume and direct traffic growth. Treat it as a poll of your buyers, not as a ledger. Polls have margins of error and they are still how we measure things that resist counting.
The one question that captures AI discovery
The common mistake is a single vague box. "How did you hear about us?" as open free text yields answers like "online" and "a friend" that nobody parses. The better pattern, recommended by both Reddireach for AI specifically and Airbridge for surveys generally, is two fields that do different jobs.
| Approach | What you ask | What you get |
|---|---|---|
| Single free-text box | "How did you hear about us?" | Rich detail in theory, unparseable mush in practice |
| Structured channel + detail | A dropdown of named channels, then "what did you ask or where did you see us?" | A clean category to count, plus a confirmable AI signal |
Give the dropdown one option per real channel: ChatGPT, Google, Perplexity or another AI tool, Reddit, YouTube, a podcast, a colleague, and Other. Then add a short text field: "What did you ask, or where did you see us mentioned?" That second field is what separates a real AI discovery from a guess. A buyer who writes "asked ChatGPT for the best options" is confirming the channel in their own words. Bucketing everything as "social" or "online" destroys exactly the signal you are trying to isolate.
If you want to know when AI changes its answer about your brand, start with a free brand check — it takes a minute.
How to build it without breaking your funnel
Build it on the form that already carries intent, not on a new popup. The demo request, the trial signup, the checkout. Here is the sequence.
Attach the question to a high-intent form a buyer already completes: demo request, trial start, or first purchase. Do not add a separate survey step that adds friction.
List one option per active channel, including ChatGPT, Perplexity, and a generic "another AI tool." Add a short free-text follow-up to confirm AI discovery in the buyer's words.
Randomize option order per respondent, keep "Other" last, and pipe the answer into a CRM field so it sits next to deal value, not in a spreadsheet nobody opens.
Read the free-text field weekly. Reclassify obvious AI mentions, watch the share trend, and report the direction rather than any single week's number.
The routing step is the one teams skip and regret. If the answer does not land on the contact or deal record in HubSpot, Salesforce, or your billing system, you cannot ever connect it to revenue, and revenue is the whole point. For the broader plumbing of capturing what cookies miss, the dark SEO funnel breakdown covers where this traffic actually goes.
When to ask: the post-conversion moment that pays off
Timing changes your response rate more than wording does. Asked at a post-conversion moment, immediately after a buyer starts a trial or completes a purchase, surveys see completion rates of 45% to 85%, up to ten times higher than the same question sent later by email, per Airbridge's benchmarks. The buyer is engaged, the decision is fresh, and recall is at its peak.
Asked at the wrong moment, the data degrades fast. A delayed email survey collects a small, self-selected slice and recency bias does the rest: people name the last thing they remember, not the thing that started them. There is a real recall advantage to AI here that you should not over-read either. Buyers remember "I asked ChatGPT" vividly because it felt like a conversation, which can inflate AI's share relative to a passive channel like display. Note that tendency, do not try to correct it out of existence, and keep the question anchored to the moment of conversion where memory is most accurate.
How to keep the data honest
Self-reported data has known failure modes, and the fix for each is mechanical. Anchoring bias favors whatever option sits at the top of the list, so randomize the order for every respondent and pin "Other" to the bottom, since people who choose it are deliberately rejecting your options. Brand familiarity is the bigger trap: a buyer who saw your ad inside a mobile game will often report "Google" or "a friend" because those are the names they know. That is why reliability drops as your channel count rises.
With one or two channels, self-reported answers are high-confidence. With three or four, treat them as directional. Past five active paid channels, brand familiarity dominates recall and the raw percentages stop being trustworthy on their own. At that point, pair the survey with a causal test rather than reading the shares literally.
None of this makes the data useless. It makes it a poll, with the same caveats every poll carries. The discipline is to report the trend and the direction, cross-checked against a second source, instead of quoting a single channel's percentage to two decimal places as if a cookie had recorded it.
Reconcile what people say against what the model shows
Compare survey responses with your AI visibility results. The survey is the downstream signal: buyers telling you AI sent them. Your model-level visibility is the upstream signal: how often AI actually recommends you for the prompts that matter. When both move together, you have a credible story. When the survey shows rising AI discovery but your visibility is flat, the buyers are likely finding you through a competitor's comparison page or a Reddit thread, not a direct recommendation, and that is a different problem to solve.
Use the survey to correct your analytics, not replace them. Airbridge shows an example where a $50 reported cost per acquisition drops toward $29 once you account for the channels analytics under-credits. Apply that logic to AI: if surveys say AI influences 11% of new revenue and GA4's AI Assistant channel shows 4%, the gap is your correction factor. Feed it into the AI visibility ROI framework rather than letting the undercount stand.
Graduate to a lift test once you run several channels
Once self-reported data gets noisy past a few channels, the honest next step is a causal one. A lift test, or geo holdout, measures impact instead of asking for it. Pause or concentrate your AI-visibility investment in one segment, hold another segment steady, and compare the difference in conversions and branded demand over a fixed window. This is the measurement Fishkin points marketers toward when attribution breaks: directional lift over time, read against a control.
You will not get a clean randomized experiment for a single brand's AI visibility, because you cannot split the world's perception of you into test and control. What you get is a strong observational estimate, and that is enough to defend a budget. The sequence is a maturity ladder: start with the survey to prove the channel exists, reconcile it against model visibility to size it, and run a lift test to confirm it before you scale spend. For how AI-referred buyers behave once they arrive, the AI search conversion data is the companion read.
Who should set this up now
If you sell anything with a considered purchase and a form in front of it, B2B software, professional services, high-consideration ecommerce, you should add the question this week. It costs one field and a CRM mapping, and it starts compounding immediately because attribution data is only useful as a trend. The teams that wait are the ones who, six months from now, still cannot answer the CFO's question with anything better than a shrug.
The reader who should not over-invest is the one running a pure-impulse, low-consideration funnel where buyers cannot articulate a discovery path. There, lean harder on lift tests and visibility tracking. For everyone in between, the survey is the cheapest honest signal available, and when you walk it into a leadership review, present it as a triangulated read, not a single number. The CEO reporting guide covers how to frame a directional metric without losing the room.
