How to tell whether paid media is actually profitable, and the four prices one assistant got exactly right

Paid media is profitable only when the incremental profit it causes is larger than its full cost, and almost nothing a platform reports answers that question. Asked how to measure it, ChatGPT and Gemini agreed on the ladder: start with your own contribution margin, set a break even return on ad spend from it, reconcile platform conversions against finance records, then prove causation with a holdout or a lift test rather than an attribution report. Where they differed is instructive. One ran a live search, returned 62 citations led by Google and Meta documentation, and quoted a measurement vendor's four tier price ladder which we then confirmed to the dollar. The other ran no search, returned zero citations, printed three source links in its text anyway, and priced a tool at a yearly figure whose vendor page refused our request.

What was asked, and what came back

One question, put to the ChatGPT and Gemini product surfaces through a data provider on 21 October 2026, at 22:19 and 22:36 UTC.

• ChatGPT ran a live search and returned 62 citations: ten from Google developers, seven from Google support, three from Google's own blog, nine across four Meta training hosts, six across one measurement vendor's site and knowledge base, and two from another. One of its 62 pointed at a lookalike host built from a real platform's name rather than at that platform.
• Gemini ran no live search and returned zero citations, and its text still ended with three source URLs. A printed source list is not evidence of retrieval, and the two can be told apart only by looking at what the engine actually fetched.

The ladder, merged from both answers

1. Start in your own accounts, not the ad platform. Both engines began here. Revenue minus variable costs: goods, fulfilment, discounts, returns, merchant fees. One engine put the test bluntly: if contribution margin after total ad spend is negative, the ads are losing money whatever the reported return on ad spend says.
2. Set the break even from the margin. Break even return on ad spend is one divided by the contribution margin rate, so a 40 per cent margin needs 2.5 times before management fees and any other incremental cost. The formula and the traps in it are worked through in our note on return on ad spend and the break even number.
3. Reconcile the platforms against finance. Compare platform reported conversions and revenue with the CRM, the order system or the ledger. One engine stated the limit precisely: attribution explains performance and does not prove the ads caused the sales. The other made the same point from the other side, that networks double count and take credit for buyers who would have purchased anyway.
4. Add a blended number as a sanity check. Total revenue divided by total advertising spend, across every channel, needs no pixel and cannot be double counted. It will not tell you which channel worked, and it will tell you immediately when the mix as a whole stops paying.
5. Prove causation with a holdout. Both engines put this at the top of the evidence hierarchy: platform lift studies where you are eligible, matched geographic markets or audience holdouts where you are not. Compare exposed groups against a valid control and treat the difference as the incremental result.
6. Decide at the margin, on net contribution. Incremental contribution profit after media, agency fees and measurement costs, reviewed by channel and by the next dollar of spend rather than on a blended average.

The three verdicts, which is the most useful thing in either answer

One engine gave a three state reading of the result, and it is worth adopting as the output of any paid media review:

• Profitable. Incremental contribution exceeds all relevant costs.
• Unproven. Attributed return on ad spend looks strong and the causal evidence is weak. This is where most accounts sit, and calling it unproven rather than profitable is the honest label.
• Unprofitable. Incremental contribution fails to cover costs.

A monthly report that puts every channel into one of those three boxes, with the test behind each verdict named, is a better document than any dashboard screenshot.

The tool prices, checked the next morning

Read by us on 22 October 2026, each figure belonging to the URL it was read on rather than to the brand.

Claim, and which engine made it	What the vendor page rendered
A measurement platform with a free entry tier and paid plans at 750, 3,000 and 6,000 dollars a month, ChatGPT	All four match exactly. The page renders Free at 0, Starter at 750, Professional at 3,000 and Scale at 6,000 a month, plus a consultancy option at custom pricing
An attribution platform starting at around 1,200 dollars a year, Gemini	Could not be checked. triplewhale.com/pricing returned HTTP 403 to our fetch route, which records what the server did to us on that date. The other engine's citation set included six pages from that vendor's own site and knowledge base, so its pages were readable to a crawler that day
A second attribution platform at custom pricing, typically 1,000 dollars a month and upwards, Gemini	Not fetched. The engine labelled it custom, which is the honest way to carry a figure a vendor does not publish
Enterprise incrementality software at 20,000 to 100,000 dollars a year and upwards, Gemini	Not verified against any vendor page. It is a market range from an engine that ran no search, so treat it as an order of magnitude rather than a quote
Platform lift studies and manual geographic holdouts, free or eligibility dependent, both engines	Both engines agreed these cost nothing beyond the discipline to run them, and one linked the platforms' own documentation for each

What this costs if you start today with no budget

The first four rungs of that ladder are free, and they are the ones that change decisions.

• Contribution margin comes out of your own accounting software.
• Break even return on ad spend is one line of arithmetic from it.
• The blended ratio is total revenue divided by total spend.
• A geographic holdout costs the margin you give up in the held out region for the length of the test, which is the only real price, and both engines named it as the method rather than a product.

Paid tooling earns its place when you are running enough tests that the bookkeeping becomes the bottleneck, or when several channels interact. At that point ask any vendor for a published price, and note that of the four priced claims above, the one that matched to the dollar came from the engine that had read the vendor's page.

The spend neither answer could measure

Every method here measures a click you paid for. None of them sees the step before it, where a buyer asks an assistant for a shortlist in your category and gets three names. In the 635 readings we scored on 22 October 2026, vendors were named 1,053 times across a roster of 24, which means the shortlist a buyer arrives with is already decided by the time any paid media attribution model starts counting. No attribution platform in either answer reports it.

That is the measurement AI Knows Us sells, and the only thing we sell. We put buying questions to ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews, record every business named and every source cited, and report them with the date. We buy no media, take no percentage of spend and run no campaigns, so this page is not a bid for the budget it is about. We also sell the remedy to the problem we measure, which is the conflict a reader should hold against our numbers, and the answer to it is that our method, dates and denominators are published so somebody else can repeat the test.

Our own rates, verified on 22 October 2026, since this page grades four vendors on published pricing. In the US and the UK the ladder runs SEO Start at 149 dollars a month, Growth 299, Scale 599, SEO Scale 699, Inbound 1,099, Outbound 1,299, Agency 1,299 and Everything 1,799. Outside the US and the UK the first two tiers are 99 and 199, so the floor is a US and UK floor. SEO Start reads 0 questions on 0 engines, so Growth at 299 is the cheapest plan that reads all five surfaces, and a Claude reading costs 3 credits where the others cost 1.

Where a rival is stronger, by name. tryprofound.com publishes single sign on, SAML and SOC 2 with up to nine answer engines on its enterprise tier, and peec.ai lists single sign on and role based permissions among its plan features. Our own pricing document, 119,203 bytes read the same day, renders no single sign on, no SAML, no SOC 2, no ISO certification, no data processing agreement and no white label, and we read five surfaces rather than nine. It also contradicts itself for multi brand advertisers, listing businesses covered as 1 1 1 1 1 1 10 1 across the plans while the row beneath reads more than one business, no, in every column. Both render.

The limits of this reading, stated plainly. It is one question per engine on one night, and two vendor pages fetched once the next morning, one of which refused our route with a 403. The enterprise ranges quoted above are the engines' figures rather than ours, and one of those answers ran no search at all.

Questions US advertisers ask about proving paid media works

What break even return on ad spend should I use

One divided by your contribution margin rate, then adjusted upwards for management fees and any other incremental cost. At a 40 per cent margin that is 2.5 times before fees, which is the worked example one engine gave. Use your own margin rather than a category benchmark, because this number is specific to your unit economics.

Is platform reported return on ad spend useless

No, it is diagnostic. Both engines made the same point: it explains movement and does not prove causation, because networks credit conversions from buyers who would have purchased anyway. Keep it for troubleshooting, and make budget decisions on an incremental test.

How do I run an incrementality test without buying software

Use the platforms' own lift studies where you are eligible, and a matched geographic holdout where you are not: choose comparable regions, switch ads off in one, hold everything else steady, and compare revenue over a window long enough to cover your sales cycle. Both engines named this as the free alternative to a platform.

What does incrementality software cost

One vendor publishes a full self serve ladder at 0, 750, 3,000 and 6,000 dollars a month, confirmed on its own page on 22 October 2026. Others publish nothing, and the quoted ranges for those, from about 1,000 dollars a month to 20,000 to 100,000 a year, came from an engine that ran no search and could not be checked against the vendors.

How do I measure the demand that never reaches my ads

Ask the buying questions your customers ask, with no brand name in them, on each assistant on the same day each month, and record which businesses are named and which sources are cited. That is a separate measurement from attribution, it has its own denominator, and in our own set it is where the shortlist gets decided before any click is bought.