Can I trust “How did you hear about us?”

1. Blog
2. Measurement & Budget

By Saroj Jha · November 2, 2026 · 14 min read

You can trust it for what it is: buyers’ memory of what brought them to you. That memory sees what click tracking cannot, such as a colleague’s recommendation, a podcast, a private community or an answer from an AI assistant, and on those channels it is often the only evidence you have. It is also biased: people forget early touches, name the most recent or memorable one and lean towards answers that sound considered. In my view the answer is to ask with an open-text field at the moment of conversion, code the answers with a written codebook that two people can apply the same way, read them by quarter rather than month, and act only when self-reported and software attribution agree, or when a test settles the disagreement.

Here is what the question catches, where it misleads and how to use it.

Running case: Example Co. Fictional

Example Co. is a fictional seed-stage B2B SaaS company selling an AI contract-review tool to in-house legal teams at companies with 100 to 1,000 employees. Its founder posts on LinkedIn, it sponsors one legal-operations podcast for $9,000 a quarter and it spends $24,000 a quarter on paid search and social ads. Of last quarter’s 240 demo requests, its CRM records the last-click source of each, and its demo form asks “How did you hear about us?” in a free-text box. Example Co. and every number attached to it are illustrative.

What does “How did you hear about us?” catch that click tracking misses?

Everything that happens where no link is clicked. A colleague who says your name over lunch, a podcast host who mentions you, a thread in a private Slack community and a screenshot passed around a team leave no referrer. When the buyer later types your address or searches your name, analytics records “direct” or “organic search”, and the channel that did the work gets nothing. I cover why so much B2B demand now starts in those places in Dark Social: Where B2B Demand Actually Happens Now, including a vendor test of software against self-reported attribution, so I do not repeat it here.

AI assistants add a newer gap, because a buyer an assistant sends your way may never click a link to your site at all. Pew Research Center’s analysis of browsing data from 900 US adults in March 2025 found that Google users who met an AI summary clicked a traditional search result in 8% of visits, against 15% for those who did not, and clicked a link inside the summary in 1% of visits^[2]. That study is about Google’s AI summaries, not chat assistants, but the direction is the same: more of the research happens on a page you never see. Self-reported attribution is one of the few ways to hear about it, and the AI Referral Share Reader shows how much of the share you can still see in your own traffic.

Example Co., step 1 Fictional

Example Co.’s CRM credits its 240 demo requests to direct (96, 40.0%), organic search (70, 29.2%), paid search and social (30, 12.5%), LinkedIn (24, 10.0%), referral (12, 5.0%) and AI referral (8, 3.3%). In their own words, the same 240 named a colleague or peer 62 times, Google search 41, the founder’s LinkedIn posts 38, the podcast 31, an AI assistant 22, a community 19, an event 9, a paid ad 8, and 10 answers were blank or unclear. Peers, the podcast, AI assistants, communities and events usually leave no referrer, and together they account for 62 + 31 + 22 + 19 + 9 = 143 answers, or 59.6%.

Do this

Put last quarter’s tracked sources and buyers’ answers side by side for the same conversions. Every source buyers name that has no tracked equivalent is a channel your dashboard cannot see.

Why can’t you take the answers at face value?

Because the answer is a memory, and memory has known habits. Survey researchers have studied them for decades. Sudman and Bradburn’s review of response effects in surveys names “The four most important variables relevant to memory” as recency, importance, complexity and affect^[4]. A source a buyer met recently, or that mattered to them, is easier to recall than the blog post they read eight months ago.

Three biases follow, and each pushes answers in a predictable direction:

• Recall bias. Early, low-stakes touches fade, so first exposures are under-reported. A buyer who first saw you in a newsletter a year ago may honestly not remember it.
• Last-salient-touch. People tend to name the most memorable recent step, which is often the one that prompted the visit rather than the one that created the interest. “Google” often means “I searched your name after someone mentioned you.”
• Social desirability. People give answers that present them well. A meta-analysis of list experiments found that sensitivity biases are “typically smaller than 10 percentage points and in some domains are approximately zero”^[5]. Where a source sits on a demo form is not a sensitive topic, so I expect this bias to be small, but in my view it tilts answers towards sources that sound considered (a colleague, an article) and away from ones that sound impulsive (an ad).

Self-reports of media use in general are rough. A 2021 meta-analysis in Nature Human Behaviour, based on 106 effect sizes, found that “self-reported media use correlates only moderately with logged measurements” and “self-reports were rarely an accurate reflection of logged media use”^[6]. That study measured time on screens, not attribution, but it is a fair warning: treat each answer as evidence of what the buyer remembers, not a log of what happened.

Example Co., step 2 Fictional

Of the 41 buyers who wrote “Google”, the sales team asked 20 on the first call what made them search. Eleven said a colleague or the podcast had mentioned Example Co. first. That is a small, unrandomized check, so Example Co. does not recode the 41 answers; it records that “Google” partly hides earlier word of mouth and asks a follow-up question in future (next section).

Do this

For each bias, write down which of your channels it probably flatters (search and recent campaigns) and which it probably hides (early content and newsletters), and keep the note beside every report.

Should you ask with a dropdown or an open text box?

Open text, because a list changes the answers. Pew Research Center’s methods guide describes a question asked after the 2008 US presidential election: “What one issue mattered most to you in deciding how you voted for president?” When the economy was offered as one of five options, 58% chose it; when people answered in their own words, 35% volunteered it^[1]. Among those given the list, 8% gave an answer outside it, while 43% of those answering freely gave an answer the list did not include^[1].

A source dropdown does the same thing to attribution. Buyers pick the closest option, and channels you did not think to list, a particular podcast, a Reddit thread or an AI assistant, disappear into “Other” or into “Search”. Order matters too: Pew notes that in self-administered surveys respondents “tend to choose items at the top of the list”^[1]. Open text costs you coding time, and it gives you the buyer’s own words, including sources you did not know existed.

Example Co., step 3 Fictional

Example Co.’s form used to offer a six-option dropdown (Google, LinkedIn, Event, Referral, Ad, Other). It switched to a free-text box with the prompt “How did you hear about us? A name, show or tool helps” and a second optional box: “If you searched for us, what made you search?” The podcast and AI assistants, which the dropdown had no option for, now appear by name.

Do this

Replace any source dropdown with one open-text box, plus an optional follow-up for people who answer “Google” or “search”.

Where in the flow should you ask?

At the conversion that matters, as close to the decision as you can. For a sales-led company that is the demo request; for a self-serve product, the signup. The longer you wait, the more the answer reflects recent touches, including your own onboarding emails, rather than what brought the buyer in.

Make the field required only if it does not cost you conversions you care about. My rule of thumb, not a benchmark: required on a demo form, where the buyer has already decided to talk to you, and optional on a free signup, where every field costs completions. Ask again on the first sales call for high-value deals, because a conversation gets a fuller answer than a form, and record that answer in the same field so it can be coded with the rest. Do not ask on newsletter signups or content downloads; you want the answer from people who became pipeline.

Example Co., step 4 Fictional

Example Co. asks on the demo form (required) and again on the first call, where the account executive asks “Before you filled in the form, where had you come across us?” and pastes the answer into the CRM. The 10 blank or unclear answers on the form are filled from the call where one happened.

Do this

Check where your question sits today. Move it to the demo or signup step, and add it to the first-call notes template for your larger deals.

How do you code the answers?

With a short written codebook and two people who test it. Free text is only useful once each answer has a code, and the codes are only useful if two people would assign the same one. Start with eight to twelve codes that match the decisions you make: one per channel you fund or could fund, plus “peer or colleague”, “AI assistant” and “unclear”. Keep the original text beside the code so you can recode later, and add a second field for the specific source (which podcast, which community) where buyers name one.

Then check that the codebook works: have two people code the same sample independently and measure agreement. McHugh’s methods paper on interrater reliability notes that “many texts recommend 80% agreement as the minimum acceptable” and argues that “any kappa below 0.60 indicates inadequate agreement”^[3]. Kappa corrects percent agreement for agreement that would happen by chance.

If you fall short, the codes are ambiguous; tighten the definitions, not the coders. The Self-Reported Attribution Coder applies a codebook to your answers and flags the ones it cannot place.

Read the shares by quarter, with intervals. A monthly count of a few dozen answers swings with chance, so small differences between channels mean little.

Example Co., step 5 Fictional

Two people code the same 60 answers into five groups. They agree on 51, which is 85%. Chance agreement, from how often each coder used each code, is 20.6%, so kappa is (0.850 − 0.206) ÷ (1 − 0.206) = 0.81, above McHugh’s 0.60 line. For the quarter, AI assistants were named by 22 of 240 buyers: 9.2%, with a 95% interval of 6.1% to 13.5%. In a single month of 80 answers, 7 mentions would be 8.8% with an interval of 4.3% to 17.0%, too wide to tell a trend.

Do this

Write your codebook on one page, have two people code the same 50 to 60 answers, and compute agreement and kappa before you report any share.

How do you use it alongside software attribution and tests?

As one of three views, each with a different blind spot.

Software attribution is cheap and daily but sees only clicks, and it over-credits the last click, branded search above all. Self-reported attribution sees the private channels but depends on memory. An incrementality test, such as a holdout or geo test, measures cause, but it is slow and covers one channel at a time. My rule of thumb, not a benchmark: act on a channel when two of the three views agree, and run a test when software and buyers disagree about a channel large enough to matter.

The disagreements are where the value is. A channel that software credits heavily and buyers rarely name is a candidate for a test before you scale it, because it may be collecting credit for demand created elsewhere. A channel that buyers name often and software never sees is a candidate to protect from cuts, and to measure some other way. Incrementality Testing for Startups shows how to run the test, and the tracking setup guide keeps the software view clean.

Example Co., step 6 Fictional

Paid search and social: software credits 30 demos, buyers name ads 8 times. On $24,000 of spend that is $24,000 ÷ 30 = $800 per demo on the software count and $24,000 ÷ 8 = $3,000 on the buyers’ count. The views disagree on a channel worth testing, so Example Co. plans a branded-search holdout next quarter before changing the budget. The podcast: software sees nothing, buyers name it 31 times, and the quarter’s $9,000 sponsorship works out at $9,000 ÷ 31 = about $290 per mention. Example Co. keeps it and adds a show-specific landing page so part of its effect becomes visible to software too.

Do this

Build a three-column table, one row per channel: what software credits, what buyers say, and whether a test exists. Mark every row where the first two disagree by more than half, and pick the largest for a test.

What changes by stage?

• Pre-seed — Ask every early customer on a call how they found you and write the answer down word for word. With a few dozen answers, read them all rather than coding them.
• Seed — Add the open-text field to the demo or signup form, write a codebook, check agreement once, and review shares quarterly beside software attribution.
• Series A–B — Code every answer, report shares with intervals, split them by segment, and use disagreements between the three views to choose which channel to test each quarter.

Free tool

Code buyers’ answers into channels

The Self-Reported Attribution Coder turns a column of free-text answers into coded channels and specific sources, and lists the answers it could not place so you can tighten your codebook.

Self-reported attribution is one input to the channel decisions in Marketing Analytics for Startups: Measure What Changes a Decision, which sets out how often to review it and what else sits beside it. For the channels it most often reveals, start with Dark Social: Where B2B Demand Actually Happens Now and Podcasts for B2B Startups.

If you want me to look at your own funnel, positioning and channels, my Growth Audit ends with three prioritized fixes and a 30-day action plan.