What tracked attribution cannot see
Every attribution model, whether a simple rule or a data-driven estimate, works from the touchpointsAny interaction between a person and a brand, paid, owned or earned, before, during or after a purchase. that were recorded on a person's path. A touchpoint is recorded when a visit arrives with something that identifies its source: campaign parameters on the link, which Google's URL builder documentation describes as the way to collect campaign data from custom URLs, an advertising click identifier, or a referring site the browser passed on. The attribution models comparison explains how each rule then divides credit among the recorded touches. None of them can divide credit among touches that were never recorded.
The unrecorded touches fall into a few recognisable groups.
- Word of mouth. A colleague, a friend, an accountant or a supplier mentions the business. The person later searches for the name or types the address, and the session is credited to brand search or direct.
- Audio and video without links. Podcasts, radio, television and video where the call to action is a name to remember rather than a link to tap. The guide to podcasts and audio in the marketing mix explains why audio creative leans on a memorable name, code or short address for exactly this reason, and the guide to measuring YouTube's contribution to the business shows how answers naming a channel's videos sit beside tagged visits and branded search in one report.
- Communities and private sharing. Links pasted into messaging apps, email threads, private groups and workplace chat tools usually arrive without a referrer, which is why this traffic is often called dark social and why it inflates direct.
- Offline media and events. Print, outdoor, conference talks, stands and printed material, where any link has to be typed or scanned and most people simply search later.
- Content consumed long before the visit. An article read months earlier, outside any lookback window the analytics tool applies.
Each of these hands its credit to whatever channel captured the demand, most often branded searchSearch queries that contain a company's or product's brand name., direct and retargeting. The diagnosis of retargeting and brand search taking all the credit covers the symptom in reporting; this guide covers the cheapest way to see behind it, which is to ask.
Wording the question
The wording decides what the answers can tell you, and two choices matter more than the rest.
Open text or a list. A list is easier to analyse and worse at its job. Its options are the channels the team already believes in, so it can only confirm the existing picture; the person who heard about you from a podcast the team has never sponsored has no option to choose. Lists also lead the answer. People pick the option that looks closest, pick the first plausible one, or pick the channel they think is expected, and a list that includes "search engine" will collect many answers from people who searched for a name they heard elsewhere. An open text field asks people to recall rather than recognise, which is harder and more honest. Its cost is coding, covered below, and that cost is the reason the method works.
A hybrid is possible: a short list followed by a required "tell us more" field, or an open field with a placeholder that suggests the kind of detail you want without naming channels. If you use a list, keep its order fixed or rotate it deliberately, include "a person recommended you" and "other", and still read the free text before trusting the list.
What exactly you are asking. "How did you hear about us?" asks about first exposure. "What made you get in touch today?" asks about the trigger, which is often a search or an email. Both are useful, and they are different questions with different answers. Pick the one that matches the gap you want to fill, which for most businesses is first exposure, and keep the wording identical over time. A rewording is a break in the series, and it belongs in the measurement log with its date.
Where and when to ask
The placement decides who answers and how fresh their memory is.
| Placement | Who answers | Strength | Watch for |
|---|---|---|---|
| Lead or demo form | Prospects at the moment of enquiry | High intent, answers sit on the lead record from the start | A required field adds friction; junk answers when forced |
| Checkout | Buyers at the moment of purchase | Ties answers to revenue directly | Any extra field competes with the purchase; keep it optional or move it |
| Confirmation page or post-purchase survey | Buyers shortly after purchase | No friction on the purchase itself; room for a longer answer | Fewer people answer, and those who do are not a random sample |
| Sales call or onboarding | Prospects or customers talking to a person | Richer answers, follow-up questions, names of who recommended you | Depends on the rep asking every time and recording the answer the same way |
For a business with a sales team, the form and the call can work together: the form collects the first answer, and the rep confirms or expands it and records the result in the CRMThe system that stores an organisation's contacts, companies, deals and interactions, and the processes built around it.. For a shop, the confirmation page or a short survey shortly after the order usually beats a checkout field. Whatever the placement, the answer must land on the same record as the tracked source, which means a field on the lead, contact, deal or order rather than a separate survey tool whose responses cannot be joined back. The confirmation page is also where conversion research asks buyers what nearly stopped them, and the tutorial on on-page polls and post-purchase surveys keeps that question separate from this one so that neither answer muddies the other.
Asking everyone in the same place is more important than asking in the ideal place. A question that appears on some forms and not others, or that sales asks only when it remembers, produces answers from a sample nobody can describe.
Coding the answers
Free text becomes data through a codebook: a short list of categories, each with a definition and examples of answers that belong to it. A workable first version keeps the categories at channel level, such as podcast, peer or colleague recommendation, search (unspecified), organic social, paid social, video, event or conference, review or comparison site, press, community or forum, partner or reseller, offline media, and unclear. Add a second field for the named detail, such as the podcast, the community or the event, because the detail is often what the budget decision turns on.
Five rules keep the coding usable.
- Code the meaning, not the words. "Saw you on my feed" is organic or paid social depending on what else the answer says; "Google" is search unless the answer says it was an advert. When the answer cannot tell you, use the unspecified category rather than guessing.
- One primary code per answer, with an optional secondary code when someone names two sources. Analyses run on the primary code; the secondary code shows combinations.
- Keep the scheme stable. Add a category when a new source appears repeatedly, record the date it was added, and never recode history silently. A category split halfway through the year makes the earlier months incomparable unless they are recoded under the new rules.
- Check consistency. When more than one person codes, have them code the same sample independently now and then and compare. Disagreements show where the definitions need examples.
- Code on a cadence. Monthly is enough for most volumes. Answers coded long after the fact cannot be followed up with the sales team.
An automated first pass, whether keyword rules or a classifier, can speed this up, provided a person reviews it against the codebook and the unclear pile is read rather than discarded. The unclear pile is where new sources first appear.
Free-text answers need the same care as any other customer data. People write names: the colleague who recommended you, the consultant who set up their system. The GDPR defines personal data broadly, as any information relating to an identified or identifiable person, and its data minimisation principle applies to answers as much as to form fields. Keep the raw text in the CRM, where access is managed, report the codes rather than the text, and never copy the answers into an analytics tool; the guide to keeping personal data out of analytics explains why. This is general practice, not legal advice.
Comparing answers with the tracked source
The value of the method is in the comparison, and the comparison needs both answers on one record: the self-reported code and the tracked source the system recorded for the same lead or order. For leads, that usually means hidden form fields that capture the campaign parameters and referrer at submission, a pattern the UTM and naming conventions guide depends on, so that the tracked channel is stored on the lead rather than inferred from analytics later.
Build a cross-tabulation with the tracked channel in rows and the self-reported category in columns, counted over a full month or quarter. Then read it in three directions.
- Along the diagonal. Where the two agree, as when paid social leads say they saw an advert on social, the tracked model and the person tell the same story.
- Down the columns. Take each self-reported category and look at which tracked channels its leads were credited to. Leads who name a podcast and were tracked as direct or brand search show where the podcast's effect is being credited.
- Across the rows. Take each tracked channel and look at what its leads said. A large share of brand search leads naming a colleague or an event shows a channel capturing demand that was created elsewhere, the same pattern assisted conversions hint at within the recorded paths.
| Pattern in the cross-tab | Likely reading | Next question |
|---|---|---|
| Direct and brand search leads mostly name one unrecorded source | That source is creating demand that tracking credits to capture channels | Can the source be tested, for example by pausing or varying it? |
| A paid channel's leads mostly name other sources | The channel is often the last step for people who were already coming | How much of its credit would remain in a holdout? |
| A channel is named often but tracked rarely | Its effect arrives through other routes | Is the activity measurable by codes, vanity addresses or a regional test? |
| A channel is tracked often but rarely named | It may be forgettable rather than ineffective, or it may be capturing | What does a controlled test say? |
| The unclear share rises | The question, the placement or the coding has drifted | What changed in the form or the codebook? |
Read the gap by channel, not as a single overall figure. The overall agreement between tracked and reported sources says little; the channels where they disagree most are where the budget conversation should start.
Biases and how to read around them
Self-reported answers are shaped by how memory and surveys work, and each bias has a direction.
- Recency. People remember the last thing that prompted them more easily than the first thing that introduced them, which pulls answers towards search and recent adverts.
- Memorability. A recommendation from a person or a voice on a podcast is remembered; a display banner seen in passing is not. Vivid channels are over-named and ambient ones under-named.
- Prompting. Any list, example or placeholder text shifts answers towards what it mentions.
- Who answers. With an optional question, the people who answer are more engaged than those who skip it, and their sources may differ.
- Naming confusion. People name the platform they used rather than the channel: "Google" for any search, ads included; a social network's name for organic posts and paid placements alike.
- Social desirability. "A friend recommended you" can feel like a better answer than "I clicked an ad".
None of these can be removed, but most can be held constant. Ask the same question, in the same place, with the same coding, and read changes over time rather than the level in any one month. A channel's share of answers rising after its activity increased is better evidence than its share in a single quarter. Compare self-reported trends with independent signals that move for the same reason, such as branded search volume after a podcast campaign, and treat agreement between them as corroboration.
Using it in budget decisions
Self-reported attribution is evidence about where attention came from, not about what a channel caused. A podcast named by many buyers may have reached people who were going to buy anyway; a channel nobody names may still be doing quiet work. That is why the answers should decide what to test rather than what to spend. The incrementality testing guide covers the designs that measure cause; this method points them at the right questions.
In practice the answers earn three uses. They add a view beside the tracked models in reporting: the business-to-business measurement guide reports sourced and influenced pipeline with self-reported attribution as a third view, and the same layout suits any business with a long path. They prioritise tests, pointing a holdout at the capture channel that self-reports suggest is over-credited or at the unrecorded channel that seems to be under-credited. And they explain disagreements, giving the choosing an attribution approach guide a concrete input when a channel's tracked credit and its perceived importance diverge.
Wrenmarsh Payroll adds an open question to its demo form
Wrenmarsh Payroll, an invented payroll software firm, adds a required open field, "How did you first hear about us?", to its demo request form and stores the tracked source on the same record through hidden fields. After one quarter it has 418 coded answers. Tracking credits 147 of those leads to paid search, 121 to direct, 64 to organic search, 52 to paid social and 34 to referrals. The coded answers tell a different story: 112 name a podcast, 96 name a colleague or their accountant, 71 name search, 58 name a social post or advert, 37 name an event and 44 are unclear. Of the 121 leads tracked as direct, 49 name a podcast and 38 a peer recommendation, and of the 147 tracked to paid search, 61 say they first heard about the firm on a podcast or from a colleague. The team does not move budget on this. It books a regional holdout on paid brand search for the next quarter to see how much of that channel's credit is demand created elsewhere, and adds a sponsor code to the two podcast reads so that the next quarter's answers can be checked against a second signal.
The attribution hub places self-reported attribution beside rule-based models, experiments and mix modelling, each answering a different question about the same customers.
Common mistakes
- Using a dropdown. It can only confirm the channels the team already listed.
- Asking in some places and not others. The answers then describe an unknown subset of customers.
- Storing answers where they cannot be joined. Without the tracked source on the same record there is no comparison, only a pie chart.
- Changing the wording or the codebook without a date. The trend breaks and nobody knows when.
- Reading the level instead of the trend. Biases are stable enough to compare months, not to read one month as truth.
- Moving budget on answers alone. Being named is not the same as causing the purchase; test first.
- Copying free text into analytics or reports. Answers contain names; report codes and keep the text in the CRM.
- Ignoring the unclear pile. New sources show up there first.
