How the bench will test AI ad tools

What a research edition rests on

A research edition rests on documents you can open yourself. Most are the vendor’s own: its live pricing page, help centre and terms. Some pages also count dated user reports. Every claim sits on one rung of the evidence ladder below, and its wording shows which one.

An example of each rung: editorial policy and corrections.

A price is verified when it was read on the date shown, on the vendor’s live pricing page, official help centre or developer documentation. A price we could not read there shows “Unverified”, and a vendor with no public price list shows “Not published”.

When this page was built, 59 of the 76 tools in the database had a verified price and 15 showed “Unverified”. Of the rest, 1 had no public price list and 1 had shut down.

A terms summary counts as verified only when the clause itself was read in the vendor’s terms. A plan card does not count.

How often prices and terms are re-checked

Prices are re-checked against the live official page at least every 30 days, and terms at least every 90 days. After 30 days a price stops counting as verified. Tables then replace the figure with “Unverified, last checked” and the date; sentences keep the figure beside the date it was last read.

Each re-check moves the checked date in a page’s meta line; “Updated” moves only when a fact changes, which the page’s list of changes records. The price tracker dates every tool’s prices.

The AI commercial use matrix does the same for terms: every row shows the date its clause was read.

What the bench status labels mean

Every tool page carries one of four bench status labels:

• “Not yet bench-tested”: the tool has not run the brief, so its pages are research editions.
• “Bench-tested” with a date: the tool ran the brief on that date, and its outputs are published.
• “Re-test due”: the tool ran the brief, but a new model or new pricing since then means its results may no longer hold.
• “Discontinued”: the tool has shut down, and its page says what happened and what replaces it.

On 24 September 2026 no tool in the database had run the brief, so every tool reads “Not yet bench-tested”, or “Discontinued” where it has shut down.

The fixed brief: what is decided and what is not

The bench brief will be one product, one hook line and a fixed set of ad formats, identical for every tool. Only the tool then changes between two sets of outputs. None of the three is chosen yet.

What the bench will measure and publish

Every output a tool makes on the brief will be published, failures included, next to the brief that produced it. The three measures below are proposals. Who judges an output usable, and against what checklist, is not decided.

Research editions quote a different figure today: the cost per output at list price. It divides a plan, top-up or API price by what that price buys, then multiplies by what one output uses.

Each tool’s method note shows its own sum. Where a plan counts credits, the figure holds only if every credit is used and nothing is re-rendered.

Tools whose terms keep them off the bench

The terms below bar benchmarking the service or publishing the results. Meta’s go further and bar any use of its output off its own platforms.

On our reading, each benchmark clause covers running a tool on a shared brief and publishing the comparison. None of these tools will run on the bench without its vendor’s written agreement, so their pages stay research editions.

Motion makes no ad creative to bench anyway.

Canva’s live terms page refused our requests on 24 September 2026. Its clause was read in the Wayback Machine capture of 23 September 2026, which carries the same effective date. We have not read whether Canva’s Enterprise agreement says the same.

Jasper’s clause does not apply to a use expressly permitted, for example in its Documentation or an Order Form.

Meta’s clause names no benchmark. Its ban still keeps every Advantage+ creative output off the bench and off every page here. Its words come from the Wayback capture of 22 November 2024; the live page still showed the 6 May 2024 date on 24 September 2026.

Vidu’s clause is unclear. Its Terms of Service, section 5, last updated 3 July 2026, name no benchmark but bar “Using the Services for competitive analysis, developing competing products or services, or any purpose that may be detrimental to our business interests”.

On our reading a published side-by-side comparison may count, so Vidu stays off the bench unless it agrees in writing.

Before each round, every tool’s terms will be re-read for benchmark clauses, and this section updated to match.

How tools enter the database, and what happens when they leave

A tool enters the database when it makes ad creatives, edits them or helps test them. That covers AI UGC and avatar video, generated video, product video, static ads, ad copy and creative research. No vendor can pay for a listing or a better place.

A tool that shuts down or leaves ads keeps its one page, with no child pages or comparisons. Which tools run in which round is not decided, and will be published here with the brief.

How we stay independent

Every bench account will be bought at the price checkout charges any buyer, with the list price recorded beside it. List price is a plan’s regular price on its official pricing page, before any coupon, partner deal or limited-time offer. Gifted plans and credits will be refused. No vendor will see an output or draft before it is public.

Verdicts are written without commission data, and each verdict’s wording is recorded before anyone checks which linked tools pay us. Who pays us lists every tool in the site’s database and whether it pays. A correction is dated and stays visible on the page it fixes; the about page says how to report one.

Method versions

• Draft 0.3, 24 September 2026: Canva, Jasper and Meta Advantage+ creative joined the tools kept off the bench, and Vidu’s clause was added as unclear.
• Draft 0.2, 24 September 2026: the unapproved product and hook line came off, and open parts are now labelled as proposals. The benchmark clauses on file that day were quoted.
• Draft 0.1, 23 September 2026: the first draft of the method.