A podcast download is not a listener, and a listener is not a customer

Podcast advertising is sold on downloads and measured with pixels, and both are estimates about devices and households. A business that closes by phone or in person can do better: a number and a page that exist only for the show, sales matched back to the people who used them, and a timing comparison for the rest.

Podcasting has become a serious advertising medium while keeping a measurement system built for a much smaller one. In 2025, United States podcast advertising reached $2.862 billion, up 17.6% on the year, and Edison Research's annual survey found that 58% of Americans aged twelve and over had listened to a podcast in the previous month. That is an audience the size of a broadcast network, bought largely on a metric called a download.

The trouble is that a download is not a person, a person is not a listener, and a listener is not a customer. Each step loses information, and the industry's answer to the loss has been to add inference: audience estimates, pixel matching, modelled conversions. This post explains what each layer actually measures, and then sets out the route a business with its own sales records can take instead.

What a podcast download actually counts

A podcast download is a server-log event. Under the IAB's Podcast Measurement Technical Guidelines, version 2.2, it counts once at least sixty seconds' worth of the audio file has been delivered to one device, deduplicated so that one listener fetching one episode counts once a day, with known bots and Apple Watch fetches removed.

That definition is a real improvement on the one it replaced, and hosting companies that follow it will say so on their invoices. It is also worth reading slowly. Sixty seconds of a file is a threshold on bytes transferred, not on audio played. The deduplication key is a combination of IP address and user agent, which is a fair proxy for a listener until a household shares a connection or a phone changes networks. And nothing in the definition can see whether the episode was opened at all.

The best evidence for how far downloads sit from listening came from Apple, by accident. In iOS 17, released in September 2023, Apple Podcasts stopped automatically downloading episodes older than seven days when a listener resumed a show they had let lapse. Nobody listened less. The count simply stopped including files that had been fetched and never played.

Every download number you have been quoted before October 2023 was inflated by files nobody played, and the industry found out because Apple stopped sending them. Treat a download as delivery, never as attention.

How pixel attribution matches a listener to a sale

Podcast attribution pixels work by matching. A script on your website records the IP address of each visitor; the attribution platform records the IP address of each device that downloaded an episode carrying your ad; a visit from an address that also downloaded the episode is attributed to the ad, within a window, usually thirty days.

Podcasting has no cookie and no click, so IP matching is the only signal available to a platform, and the largest of them is candid about what that involves. Spotify Ad Analytics, the service Spotify built out of Podsights after buying it in 2022, describes a device graph that decides which addresses belong to a household, discards connections from mobile towers and offices as too noisy to trust, and then models the results back up to account for the impressions it had to discard.

None of this is dishonest. It is the best that can be done from the platform's side of the wall, and the documentation says so more plainly than most advertisers read it. But notice what the number at the end of that chain is. It is a count of website visits from households that a graph believes also downloaded an episode, multiplied up by a ratio, within a window somebody chose. It is not a sale, and it cannot be one, because the platform never sees your sales.

The same objection applies to modelled conversions everywhere, and we have made it before about Google's and Meta's modelled figures. A modelled number is an estimate with the uncertainty removed from the presentation. The estimate may be good. You still cannot reconcile it against your books.

The three signals a podcast ad leaves behind

A podcast ad leaves three kinds of evidence, in decreasing order of certainty: direct responses to a number, code or page that exists only for the show; a household-level inference from pixel matching; and a change in total sales, branded search and inbound calls while the ad ran. Only the first can be matched to a closed sale.

The right answer is not to pick one. It is to use the exact methods for what they can prove and the comparison for what they cannot, and to keep the two numbers separate on the page so nobody adds an inference to a fact.

The route for a business that closes by phone or in person

Give each show its own way in, match what comes through it to your closed sales, and compare the whole business before and during the campaign. A tracking number and a landing page give you revenue you can defend; the comparison gives you the effect that never touched either.

1. One number and one page per show

A host reading out a phone number is the oldest response mechanism in advertising, and it still works because the listener is usually not looking at a screen. The number should appear nowhere else, for the same reason a billboard's number should not: the moment it is on your website too, the calls it receives stop meaning anything. The same goes for the page. A short path, read out on air, with nothing on it that the main site does not also have.

If you already use call tracking, a podcast number is one more source in the same export. If you do not, a single forwarding number per show is enough; you are not trying to track a website visitor's journey, only to know which show the caller heard.

2. Match the calls and the visits to closed sales

This is where a business with its own sales records has an advantage over every advertiser measured by a platform. The calls to the show's number have a phone number attached. The form on the show's page has an email attached. Your closed sales have both. Matching them, exactly, by normalised phone number and email address, tells you which callers became customers and what they were worth, in the same currency your accountant uses.

The match rate will not be 100%, and it should not be reported as if it were. Some callers give a different number from the one they buy with; some listeners hear the ad and walk in. Those sales stay unattributed rather than being spread across channels by a model, which is the rule that makes the attributed number worth quoting. We wrote about why match rates below 100% are the honest ones, and it applies here without change.

3. Compare the weeks, and the regions if you have them

The matched sales are a floor. Most of a podcast's effect arrives sideways: somebody hears the host recommend you, remembers the name, and searches for it a week later. The number never rings; the page is never visited; the sale still happened because of the show. The only way to see that effect is a comparison.

• Before and during. Total closed sales, inbound calls and branded search in the four weeks before the ad ran against the four weeks it was on air. A national show gives you no untouched region, so time is the control.
• Where the show is not. If a show's audience is concentrated, as regional or genre shows often are, the areas it does not reach are a comparison group. Design the comparison before the flight, not after.
• One show at a time. Two shows launched in the same week cannot be told apart by any comparison. Stagger them, even if it costs a fortnight.

The arithmetic is simple enough to do in the same spreadsheet as the matched sales. Suppose a roofing company closes 40 jobs in a typical four-week period, at an average of $8,500. During a four-week podcast flight it closes 47, of which 4 came through the show's number and matched to a closed job. The matched revenue is $34,000 and can be defended line by line. The lift is 7 jobs, or $59,500, of which the number saw four. The honest report says both: $34,000 attributed, and a lift consistent with roughly $60,000 that includes it, subject to whatever else changed that month.

Report the matched revenue as attributed and the lift as evidence. The first is a fact about specific sales; the second is an argument about the period. A report that presents the second as the first will be believed exactly once.

What the platforms' numbers are still good for

Download counts tell you whether the show has the audience you were sold, and pixel attribution tells you whether listeners visited your website at a rate above the background. Neither replaces matched sales, but both are useful checks on the buy itself.

A show that reports 40,000 downloads an episode under IAB 2.2 is a show that delivers 40,000 files a week to devices that follow it, which is a fair statement of reach. Ask whether the figure is IAB-certified, and ask what happened to it in October 2023: a show whose number did not move after iOS 17 either had few Apple listeners or was not counting the way the guidelines require.

Pixel attribution is a reasonable early-warning system. If the platform reports no lift in visits from matched households across a month-long flight, the ad is probably not landing, and you can learn that before your own comparison has enough weeks to say anything. Treat it as a signal about the creative, not as a revenue figure.

CloseRev matches the calls and form fills a podcast produces to your closed sales by phone number and email, from two exports, and leaves what it cannot match in Direct / Unknown rather than spreading it across channels. The number a show earned is the one you can show a client or a CFO.

How this compares with the other channels that cannot be clicked

Podcasts sit with radio, out-of-home and direct mail as media that produce customers without producing clicks. The measurement pattern is the same across all of them: an exact floor from a dedicated response mechanism, a comparison for the rest, and a refusal to let a model fill the gap between the two.

What is different about podcasts is the intimacy of the read and the precision of the audience. A host-read ad on a show about home renovation reaches people who are, by definition, thinking about their homes; a direct mail drop to the same postcode reaches everyone. That is why podcast ads convert well per listener and why the matched floor tends to be a larger share of the total effect than it is for a billboard.

It is also why the temptation to accept the platform's modelled number is stronger. The ad feels like it is working, the dashboard agrees, and the matched sales look small beside it. Resist the temptation for the same reason you would resist it anywhere: the modelled number cannot be reconciled, and a number that cannot be reconciled is a number that cannot be defended when the budget is questioned.

Buy the show for its audience, measure it with your own records, and let the platform's dashboard confirm or question the creative. A podcast that earns its keep will show it in matched sales and in a lift you designed to see; one that does not will show it there too.