Limits
What we did not check, and why
We checked one Pro account on one day, by observation. That means we did not measure search latency, did not test the refund or cancellation policy, did not verify data accuracy against the ad platforms themselves, and did not test a single competitor. This page lists every boundary of the method, and where it could be wrong.
- Document
LIMITS- Revision
A page that only lists what a method found is advertising for the method. The useful half is the other one: what the method cannot reach, what it was never pointed at, and where its results would break if an assumption underneath them turned out to be false.
Everything in the claim register rests on a narrow base — a single signed-in Pro account, observed on 4 August 2026, with no instrumentation and no money moved. That base is strong enough to settle some questions completely and not strong enough to touch others at all. Below is which is which.
We would rather a reader discount a finding of ours for a stated reason than trust one for an unstated one.
What did we actually have access to?
One account on the Pro plan, signed in, on 4 August 2026. Everything described in the register as seen in the account was seen through that one lens.
This has a consequence that runs through every record: we saw the product as a Pro subscriber sees it. We did not see it as a Starter user sees it, and we did not see it as a member of a Team plan sees it. Where a feature is gated by plan — and several are, including four of the AI categories at $50 a month each — we can report what the interface says about the gate, not what the experience is on the other side of it.
We also saw one day. A product is not a photograph, and several of the numbers we recorded had visibly moved by the second reading within that same day.
What did we not measure, as opposed to not see?
Four things, and they are the four that would need instruments rather than eyes.
- Latency. The claim that fresh ads arrive in minutes, not days is a statement about elapsed time between an ad going live on a platform and appearing in this database. Checking it means launching an advertisement of our own and timing its arrival. We confirmed that the status machinery exists and that the feeds carried current dates; we did not put a clock on anything.
- Search speed. Same category, different subject. A sub-second response claim needs repeated timed queries across different result sizes and network conditions. An impression of a page feeling fast is not a measurement, and we declined to report one as if it were.
- Deduplication quality. The product states that the same media is never counted twice, and exposes the control that does it. Whether the matching catches every re-upload — the difference between having deduplication and having good deduplication — is a property of their algorithm that no amount of clicking reveals.
- Identity stitching accuracy. Webmaster Grouping resolves domains, Page IDs, pixels and IP addresses into single profiles. We confirmed the identifier types are used. We could not test whether any given profile is correct, and this is the failure mode that would be least visible from outside: a wrongly merged or wrongly split profile looks exactly like a right one.
Did we check whether the data itself is true?
No, and this is the largest single gap in the method — larger than anything in the register.
Every check we ran was internal. We compared what the vendor says on its public pages against what the vendor's own product displays. We never took an ad marked Active in this database and went to look at whether it is, at that moment, running on Facebook or TikTok. We never took a country breakdown and tested it against an outside source.
Completeness is the same problem in the other direction. A database of a billion ads may be missing the fifty that matter to you, and there is no external vantage point from which absence can be observed.
What did we not check because it would have cost money?
The whole billing surface below the price list. We read the prices, compared the public page against the in-app page, and set the seat count to see the quoted total for a small team. We stopped there.
That leaves the promises around the money untested: the seven-day refund, the claim that cancellation is possible at any time in account settings without penalty, and the behaviour of automatic renewal. All three are policy statements. All three are exactly the kind of statement most likely to break down at the moment it matters, and the only way to check one is to trigger it.
We also did not pass through checkout, so we did not verify the note that local currency is shown at payment time, and we did not test the referral programme.
Which claims can nobody check, no matter the budget?
Three shapes of claim are unverifiable by construction rather than by our shortage of resources, and we mark them not verifiable rather than pretending a judgement.
- Claims about the whole market. The only service that does X asserts the absence of a feature across every competitor, including ones nobody has named. No finite check settles it. We apply this to any first, only or unique regardless of who it favours.
- Claims about a population we cannot enumerate. A customer count has no external register to check against, and nothing inside a customer account exposes it.
- Claims about periods longer than our observation. A monthly intake figure cannot be confirmed from a single day, and multiplying one day's counter by thirty would manufacture a number rather than verify one — intake is not uniform, and we do not know that the monthly and daily figures count the same objects.
| ID | Vendor statement | Verdict | Checked |
|---|---|---|---|
| C-18 | 1B+ new ads every month | Not verifiable | |
| C-19 | Section counters are quotable totals | Not verifiable | |
| C-20 | The only spy service grouping by webmaster | Not verifiable | |
| C-21 | Search responds in under one second | Not verifiable | |
| C-22 | 1,100+ media buyers already on board | Not verifiable |
Did we test the competitors named in the vendor's comparison table?
Not one of them. The comparison table on the vendor's landing page names competitor prices and search speeds; those numbers are one interested party's account of other people's products, gathered under conditions the table does not describe.
We record that the table exists and says what it says. We do not repeat its figures as facts about those products, and no record in the register treats a competitor's shortcoming as evidence of anything.
The same restraint applies to the third-party ratings that circulate about this whole category. A public rating is a fact about a rating page, not about a product — and where we have looked at such pages, we have found duplicated review texts and platform sanctions for manipulation often enough to distrust the genre.
Where could our own findings be wrong?
In three specific places, all of which we would want a reader to know before citing us.
- Counters are conditional and we may have quoted a state we did not fully record. Two readings on the same day gave us materially different figures for creatives and for webmasters, because a counter depends on the selected data source and the filters in force. We now attach the page state to every number, but the underlying lesson is that any single figure — ours included — is a reading, not a constant.
- A measured count is not a measure of substance. We counted 193 entries in the country filter. That is a fact about a dropdown. It is not a claim that meaningful data exists behind each of those 193 entries, and we did not test the thin end of that list.
- A contradiction may have an explanation we cannot see. Where a published figure could not be reproduced in the product, the cause may be an older snapshot or a different internal definition rather than an error. We say so in the record, mark what we observed, and resist the more dramatic reading.
How quickly does this go out of date?
Faster than a reader expects, and asymmetrically. A confirmed record tends to age well: prices and feature lists move slowly. A contradicted record can be fixed the day after we publish it, because two of the three are stale marketing copy and copy is cheap to edit.
That asymmetry means an old register drifts towards unfairness. Every record therefore carries the date of its own check rather than inheriting a single date from the page, and changes are recorded in the changelog with what changed and when.
What would change a verdict?
| Current verdict | What would move it |
|---|---|
| Contradicted | The published copy being corrected, or the product changing to match it. Either resolves the conflict, and we would record which one happened rather than quietly flipping the verdict. |
| Partly confirmed | The qualifier ceasing to be necessary — an Include / Exclude switch appearing in the Ads feed, a section count in the marketing copy matching the sidebar. |
| Not verifiable | A method appearing that we do not currently have. Instrumented timing would move the search-speed record; a month of daily readings would move the intake record. Nothing we can imagine would move a claim about the whole market. |
| Any of them | A reader showing us we were wrong. That is a legitimate route, and correcting a record in public is cheaper than defending it. |
Where exactly does the boundary run?
| Area | What we checked | What we did not check |
|---|---|---|
| Platform coverage | Which sections carry a Meta / TikTok source toggle, and what the TikTok feed returns | Whether coverage within either source is complete for any market |
| Pricing | The public price list against the in-app plans tab, and the quoted total for a five-seat team | Refunds, cancellation, automatic renewal, checkout and currency conversion |
| Data freshness | That live statuses exist as a filter and on cards, and that feeds carried current dates | Elapsed time between an ad going live and appearing in the database |
| Search performance | Nothing | Response times under any condition |
| Data accuracy | That the product's own surfaces agree with each other | Whether an ad marked Active is running on the platform right now |
| Database volume | Counters on the dashboard and in section headers, with the page state recorded | Monthly intake, historical depth, and what is missing from the corpus |
| Webmaster grouping | That four identifier types are used and surfaced in a profile | Whether any given profile is correctly merged or split |
| AI integration | The stated endpoint, authentication model, tool count and credit rules | A completed connection from an external client |
| Corporate identity | That a named legal entity, registration number and address are published | The registration itself, against a registrar |
| Competitors | That the vendor publishes a comparison table and what it contains | Any competing product, at any price, for any length of time |
Questions about the method's limits
- Why publish the gaps at all?
- Because the alternative is a page that reads as though a method has no edges, which is the single most reliable signal that it has not been examined. The boundaries also tell a reader which findings to lean on: a record from a well-covered area carries more weight than one scraped from the edge of what we could see.
- Does one account on one day mean the findings are weak?
- It means they are narrow, which is different. A single signed-in session settles some questions completely — a toggle either exists in a section or it does not, a price either matches the public page or it does not. It settles nothing about elapsed time, correctness of matching, or the state of the market. The register is written to keep those two kinds of finding visibly apart.
- Why not just test the refund and find out?
- It would be a legitimate test and we have not run it. Doing it properly means a real purchase, a real cancellation and a real refund request, documented at each step — a different scale of exercise than an observational check. Until then the billing promises are recorded as claims, not as findings.
- You mark market-wide claims unverifiable. Isn't that a dodge?
- It would be if we applied it selectively. We apply it to every claim of the form the only, the first or the unique, in either direction, including claims that would flatter a product we happen to rate well. Marking such a claim false would require the same impossible survey as confirming it.
- What should I do with a not-verifiable record when making a decision?
- Treat it as information you will have to obtain yourself, usually during a trial. Sub-second search and near-real-time delivery are both testable by any subscriber on their own queries within an hour. That is the practical value of separating the unverifiable from the unverified: it tells you exactly which questions are still yours to answer.