Attribution when you have five marketing channels and no budget

Attribution is the question of which marketing effort deserves credit for a sale. Large companies solve it with modelling, incrementality tests, and enough volume to hold experiments. You have none of those, and pretending otherwise is how small businesses end up paying for a dashboard that produces a confident answer nobody can defend.

The good news is that a business doing dozens of sales a month does not need a model. It needs a habit.

Why do attribution models fail at small business scale?

Because every model is a rule for splitting credit, and every rule is wrong in a different way.

Last click gives everything to the final touch, which systematically overpays branded search and underpays everything that created the demand. First click does the reverse. Linear splits evenly, which is tidy and has no relationship to how people decide. Data driven models need volume that most local businesses will never reach.

None of these is a discovery of the truth. They are conventions. Two vendors using two models will hand you two different pictures of the same month, both internally consistent, and there is no experiment available to you that would settle it.

There is also a growing hole underneath all of them. Google's own documentation states that traffic from its AI features arrives inside ordinary organic search traffic, with no referral parameter and no separate channel. In June 2026 Google added generative AI performance views to Search Console, and Google Search Central's announcement lists what they contain: impressions, pages, countries, devices and dates. Clicks, click-through rate, position and the user's prompt are not available, and AI Mode is not separated from AI Overviews. So one of the fastest growing surfaces in search is, by the platform operator's own description, unattributable at the click level.

The right response is not a better model. It is to stop asking the data a question it cannot answer. Ask instead which channels are clearly producing, which are clearly not, and which you cannot tell. Three buckets is usually all the resolution your decisions need.

Does anyone have real attribution data worth looking at?

One cohort does, and it is worth knowing what it found before you build anything.

Patient Prism's Dental Patient Access Report, published 2 July 2026, scored 11,552,668 patient calls across 8,280 dental practice and DSO locations in calendar year 2025. Referring source percentages come from 5,022,887 attributed calls, which is a larger attribution sample than any small business will ever assemble. 90% of attributable patient calls came from the Google ecosystem. The Google Business Profile alone accounted for 54%, organic and Google Ads together another 36%, and Meta plus print combined came in under 3%.

Grade it honestly. Patient Prism sells call intelligence software, so this is a customer cohort of practices sophisticated enough to buy call tracking, and it is one vertical. But the shape is useful. In a category where owners spend real money on social and print, the phone data says the overwhelming majority of traceable demand arrives through one company's surfaces.

The practical read for you is not "copy dentistry." It is that a channel mix argument conducted without call data is usually an argument about the smallest part of the picture.

What is the cheapest attribution method that actually works?

A question at the point of contact, asked the same way every time, logged in the same place.

"How did you hear about us?" on your booking form, at intake, or asked by whoever answers the phone. Free text, not a dropdown, because a dropdown teaches people to pick the first plausible option. Log it with the date and the value of the job.

Run it for ninety days. Then tally.

You will get three kinds of answer. Specific and verifiable, like a named referrer or a particular ad. Specific and unverifiable, like Google, which could be map pack, organic, or paid. And vague, like "I just found you," which is honest and means the person does not remember.

That third bucket is real. People genuinely do not know. A dashboard that never shows an unknown category is hiding it rather than solving it.

Compare the tally against your digital numbers. Search Console will tell you whether the Google answers are brand searches or discovery searches, which is the most useful single split available for free. Search Console is the free tool most businesses ignore and it resolves the largest ambiguity in your survey data.

What does the arithmetic actually look like?

Here is a full ninety day tally. Use your own counts.

The log. 120 completed jobs, each with an answer and a value recorded at intake.

First pass. Google is 37%derived of logged jobs and the unknown bucket is 13%derived. That unknown bucket is larger than your Instagram bucket, which is the first useful thing this table says.

Second pass, split the Google answers. Search Console for the same ninety days shows 1,000 clicks, of which 620 came from queries containing your business name. That is 62%derived brand. Apply that split to the 44 Google answers and roughly 27 of them were people who already knew your name and used Google as a phone book. Seventeen were genuine discovery.

Third pass, regroup. Referral 31, brand Google 27, and signage 18 come to 76 of 120 jobs, or 63%derived, all of which are demand you created earlier or in the physical world. Paid and social discovery is a much smaller slice than the raw Google row implied.

What changes because of this. If you were about to move budget into search ads on the strength of "Google is our biggest channel," the split just told you most of that channel is your own existing reputation being routed through a search box. The reallocation candidate is the referral engine and the sign, not the ad account.

Now check the base before you act. Twelve Instagram jobs across ninety days is four a month. A month with seven and a month with one are the same underlying rate. Telling signal from noise matters more here than anywhere else in marketing, because attribution data is noisy by construction and you will be tempted to read a two week swing as a result.

How do I decide budget when the data is ambiguous?

Change one thing at a time and give it long enough to show.

The only attribution method with real inferential power at your scale is a holdout. Turn a channel off for a defined period, watch total inquiries, turn it back on. That is a genuine experiment and it will tell you more than any dashboard.

It is also risky, so use it on the channel you most suspect is doing nothing, in your slowest season, for a defined window you commit to in advance.

Everything else is judgement applied to weak evidence, which is fine as long as you say so. Rank your channels by three things: cost, how directly you can observe the outcome, and how confident you feel. Fund the top of that list, defend the middle, and put the bottom on a review date.

What about the percentage of credit numbers vendors quote?

Treat any "this channel is X percent of the algorithm" claim as a category error, because that is what it is.

The number in circulation for local search is that Google Business Profile signals make up a fixed percentage of local ranking. It traces back to Whitespark's Local Search Ranking Factors report, which is a survey in which 47 local search experts score 187 factors. Darren Shaw, who runs it, has said in writing that the panel does not have access to the algorithm Google uses and that much of the opinion is based on observation, adding that correlation is not causation.

Two more problems sit on top of that. Secondary summaries of the report disagree with each other about the group weightings, and the weighting chart on the primary page is rendered by script, which means most people quoting a percentage have never read it off the source. A survey of expert opinion converted into a percentage of an algorithm is not attribution. It is a poll with a decimal point.

The report is genuinely useful for ranking factors in priority order. It is not a credit allocation, and a proposal that treats it as one is doing to local SEO exactly what a last click model does to your channels: converting a judgement into a number so it stops being questioned.

What should I stop trying to measure?

Some of it is genuinely unknowable and the honest move is to label it rather than buy a number for it.

Word of mouth. Someone told someone at a barbecue. It is often your largest source and it will never appear in a tool.

Brand awareness in your city. There are ways to estimate it that cost more than your marketing budget.

AI assistant mentions. These systems are non-deterministic, meaning the same question asked twice can produce different answers, so any single visibility score is an estimate presented as a fact. Kevin Indig's Growth Memo analysis adds a second problem specific to attribution: 62% of the time, when an AI system uses your content and links to it, the answer text does not say your brand name. Link tracking and mention tracking are therefore counting two different populations, and any score that adds them together is broken before it reaches your dashboard.

Offline signage, radio, and print, except through the survey answer.

Say all of that out loud in your own reporting. A report with an honest unknown column is worth more than one where everything sums to a hundred, because the second one is arithmetic rather than measurement.

You will also find your tools disagree among themselves, which is separate from attribution being hard. Website analytics and ad platforms count differently on purpose, and reconciling them is not a project worth funding.

What to do this week

Add one question to your intake, wherever first contact happens. Free text. Required if your system allows it.

Create a simple log: date, answer, job value. A spreadsheet is fine. The point is that it exists in one place and somebody owns filling it in.

Make it part of an existing routine rather than a new task. This is the same discipline as a closing procedure people actually follow: if it is attached to something that already happens at the same moment every day, it survives. If it is a separate step, it dies in week three.

Set a date ninety days out to tally it. Put it in the calendar now.

And before you spend more on any channel, check that the site can convert what you already send it. Page speed and conversion is a real relationship with a lot of myth around it, and a slow money page will make every channel look worse than it is.

Questions about your channel mix?

Email me at eric@seod.com with your channels and roughly what you spend on each per month. Five lines is enough. I will tell you which two are worth instrumenting properly, which one I would put on a holdout test first, and which one you should stop trying to measure and simply decide about.

I answer these myself and there is nothing attached to it. Rough numbers are fine.

More on measurement sits in the analytics library.