What Self Reported Attribution Sees That Click Data Cannot

What is self reported attribution?

Asking a customer where they heard about you is self reported attribution, a survey answer given at signup or checkout rather than a click recorded by a tag. The field usually reads "How did you hear about us?" and its value lands on the account record, not in the analytics property. It measures a memory, and memory is the point: it is the only instrument that can see a channel which left no referrer, no UTM string and no cookie.

Click-based attribution measures something narrower and more reliable, which is the last taggable touch before a session started. Those are two different questions and they are supposed to disagree.

	Self reported attribution	Click-based attribution
Unit of measurement	One answer per converting person	One session or one event
Population measured	Converters only	Everyone who loaded the page
Sees offline and private channels	Yes	No
Sees the actual landing page and campaign	No	Yes
Failure mode	Recall error and question design	Missing referrer and lost identifiers
Can produce a rate	No, there is no denominator	Yes

What does it measure that click tracking cannot?

It measures influence that never became a click your tag could read. Four families of traffic behave this way, and all four collapse into one bucket in an analytics report.

Google Analytics 4 defines its Direct channel by an exact rule, published in the default channel group reference: Source exactly matches "(direct)" and Medium is one of ("(not set)", "(none)"). That is not a category of intent. It is the absence of evidence. GA4 also documents a catch-all: Unassigned "is the value Analytics uses when there are no other channel rules that match the event data". Between the two, a podcast mention, a link pasted into a private Slack message, a conference talk and a colleague's verbal recommendation all report identically. The (direct) / (none) bucket in GA4 is where they land.

One newer wrinkle is documented and counterintuitive. GA4 added an AI Assistant channel for arrivals from sources like ChatGPT, Gemini, Deepseek, Copilot or Grok, matched when "the medium exactly matches ai-assistant". The same reference states that this channel "excludes Google's AI Overviews and AI Mode", which are counted under Organic Search instead. So a customer who found you through an AI answer may be filed under three different channels depending on which assistant they used, and may still show as Direct if they typed the domain afterwards. Ask them and they will say "ChatGPT".

Self reported attribution is also the only measurement that survives an attribution window expiring. A customer who first heard a podcast ad fourteen months before signing up is outside every lookback period any platform offers. They will still say "podcast".

Four private channels, one bucket

Podcast mention

Private Slack link

Conference talk

Colleague recommendation

(direct) / (none)

Why do the answers skew?

The answers skew because the field is a survey, and survey methodology has measured these effects for decades. Pew Research Center's published guidance on writing survey questions documents each of the following, with numbers.

Offering an option inflates it. After the 2008 US presidential election, Pew asked what issue mattered most in the vote, in two versions. "When explicitly offered the economy as a response, more than half of respondents (58%) chose this answer; only 35% of those who responded to the open-ended version volunteered the economy." Putting "Google" in your dropdown is the same move. It will win.

A closed list hides everything you forgot. In the same experiment, among respondents given the closed list, "fewer than one-in-ten (8%) provided a response other than the five they were read. By contrast, fully 43% of those asked the open-ended version provided a response not listed in the closed-ended version of the question." An "Other" box with a single-digit share is not evidence that your list is complete.

Position on the list decides votes. Pew states that "in telephone surveys respondents more frequently choose items heard later in a list (a recency effect), and in self-administered surveys, they tend to choose items at the top of the list (a primacy effect)". A signup form is a self-administered survey. Whatever sits first will be overcounted. Pew's own remedy is to randomise: "many sets of response options in Pew Research Center's surveys are programmed to be randomized to ensure that the options are not asked in the same order for each respondent."

How you carve the categories changes the winner. In a January 2002 Pew poll, asking whether the president should focus on domestic or foreign policy produced 52 percent for domestic and 34 percent for foreign. Narrowing "foreign policy" to "the war on terrorism" flipped it to 33 percent domestic and 52 percent for the war on terrorism. The same thing happens when "Social media" becomes "A LinkedIn post from someone I work with".

People answer to look reasonable. Pew calls this social desirability bias: "People have a natural tendency to want to be accepted and liked, and this may lead people to provide inaccurate answers to questions that deal with sensitive subjects." Its research shows respondents overstate church attendance, charitable contributions and their likelihood of voting. In a signup form, the flattering answer is "a friend recommended you" and the unflattering one is "I clicked an ad".

Check-all-that-apply degrades quality. Pew notes that "a 2019 Center study found that forced-choice questions tend to yield more accurate responses, especially for sensitive questions", and that the Center therefore generally avoids select-all lists. A multi-select "How did you hear about us" produces more total selections and less signal per selection.

One bias is specific to this field rather than to surveys generally: only converters ever see it. There is no control group, so the data cannot produce a rate. You can say Podcast outranked Reddit among people who signed up. You cannot say a podcast converts better, because nobody counted the podcast listeners who did not sign up.

How do teams reconcile it against analytics?

They join the two on a shared key and read the disagreements as a map, not as an error. The reconciliation only works if the answer is stored where it can be joined.

1. Write the answer to a field on the account or order record, alongside the identifier the analytics or CRM side already keys on. Keep the free text and the normalised category in two separate columns.
2. Capture the click-side channel on the same record at the same moment. In GA4 terms that is session_default_channel_group, which the documentation notes is evaluated on a "Paid and organic channels last click" model, distinct from the data-driven model used for the event-scoped Default channel group dimension.
3. Cross-tabulate the two. Self-reported answer down the rows, click channel across the columns. Do not average them and do not blend them into one percentage.
4. Read three cells. Self-report says Podcast and click channel says Direct: a dark channel working, invisible to the tag. Self-report says Google and click channel says Direct: a branded search or a bookmark, and the direct traffic bucket is hiding earned demand. Self-report says a source you do not run: either a mis-click on the list, or an affiliate you did not know you had.
5. Track the "Other" free text separately and re-read it every quarter. When one phrase recurs, promote it to a list option and accept that its share will jump for reasons that have nothing to do with real growth.

Two guard rails keep the numbers honest. Randomise the option order so primacy is spread across options rather than concentrated on the first one, which is Pew's own practice. And keep the field optional, because a required field converts "I do not remember" into a guess, and a guess enters your data indistinguishable from a memory.

Self reported attribution does not replace a marketing attribution model, and the two will never agree. When a paid channel looks large in the platform and small in the field, the honest reading is that the ad captured demand another channel created. That gap, not either number alone, is the finding.

Frequently asked questions

Where should the "how did you hear about us" field go?

On the signup or checkout step, after the conversion is otherwise secured, and optional. Placing it earlier costs completions; placing it after payment loses respondents who leave immediately. Pew's guidance on question order applies directly: "earlier questions can unintentionally provide context for the questions that follow", so a field that follows a list of integrations or logos will draw answers naming them.

Should the field be free text or a dropdown?

Run both, in two columns. A dropdown gives you comparable categories over time; free text is the only way to find a channel you did not anticipate. Pew documents the trade-off: a closed list drew 58 percent for an option that only 35 percent volunteered unprompted, while 43 percent of open-ended answers were absent from the closed list entirely.

Why do self reported attribution and Google Analytics disagree?

Because they count different populations with different instruments. GA4 counts sessions and assigns a channel from the referrer and UTM parameters, falling back to Direct when Source exactly matches "(direct)" and the medium is (none) or (not set). The field counts converting people and records a memory. Broader reasons that two tools report different numbers are covered in why analytics tools show different numbers.

Can you calculate a conversion rate from self reported attribution?

No. Only converters answer, so the denominator does not exist. The output is a rank order among customers and a list of channels to investigate, not a rate. Any percentage you compute from it is a share of respondents, and it should be labelled that way on every chart.

How many options should the list have?

Few. Pew advises keeping answer choices "to a relatively small number, just four or perhaps five at most", noting that people struggle to hold more in mind at once, with the exception of factual or demographic questions where longer lists help. A twenty-item list mostly measures which items sit at the top.

Does randomising the option order remove the bias?

No, it redistributes it. Pew is explicit: "Randomization of response items does not eliminate order effects, but it does ensure that this type of bias is spread randomly." Fixed order concentrates the primacy advantage on one channel permanently, which is worse, because the inflated option looks like a stable trend.

Why might someone who found you through ChatGPT show up as Direct traffic?

GA4's AI Assistant channel only matches when the medium exactly equals ai-assistant, and it covers arrivals from ChatGPT, Gemini, Deepseek, Copilot and Grok. If that same person types your domain into the browser instead of clicking a tracked link, GA4 logs a Direct session instead. Self reported attribution catches the gap anyway, because the customer just answers "ChatGPT."

What is social desirability bias in a "how did you hear about us" field?

It is Pew's term for the tendency to give an answer that makes the respondent look accepted and liked, even when it is not accurate. On a signup form, "a friend recommended you" sounds better than "I clicked an ad," so people drift toward the flattering channel. Pew's research shows the same pattern inflates reported church attendance and charitable giving.

Can self reported attribution catch a customer who converts more than a year after hearing an ad?

It can, because a memory has no lookback period the way a cookie or an ad platform does. A customer who first heard a podcast ad fourteen months before signing up sits outside every attribution window any platform offers, yet will still answer "podcast." Click-based attribution has nothing to show for that stretch of time.

What is dark social and why does self reported attribution catch it?

Dark social is traffic that leaves no referrer for a tag to log: a podcast mention, a link pasted into a private Slack message, a conference talk or a colleague's verbal recommendation. Self reported attribution catches it because it asks the converting person directly instead of reading a URL parameter that was never set. Click tracking files all four under the same (direct) / (none) or Unassigned bucket in GA4.