Methodology: how prices and claims here are verified

Every price on this site is read from a first-party source and stamped with the day it was read. When a vendor refused the request, the page names the refusal and either reads the figure from that vendor's own in-app purchase catalog or publishes nothing. This page states the rules, the sources, the gaps, and the parts of the corpus that do not yet meet them.

A methodology page is only worth writing if a reader can check it. Everything below names a specific practice, a specific source, or a specific automated check, and several sections exist to record where this site currently falls short of its own rule. The companion page, the standards that govern how articles get written, covers production, disqualifying claims, and the conflict of interest behind all of it.

What a source and a date mean on this site

A date next to a figure means the day that figure was last read from the source named beside it. It does not mean the page was rewritten, and it does not mean every other number on the page was re-checked the same day. Each figure carries its own provenance.

Two consequences follow, and both are deliberate. Pages do not restamp themselves with today's date to look fresh, because a modification date that always reads "today" is a signal that teaches search engines to ignore the field. And the sitemap emits a modification date only for pages where the site actually knows one, rather than filling the gap with the build date for everything.

Behind the pages sits a machine-readable record of every entity, attribute and value the site asserts. Each record carries the source URL it was read from, the date it was last read, and how the value was established. That last field matters more than it sounds: some records were read directly from a vendor page on a stated day, and others were established by agreement across the existing corpus and are marked as not independently re-verified. The record says which, and this page is not going to pretend the two are the same thing.

How competitor prices are checked

Prices are read from the vendor's own published surfaces, never from another comparison site. On 18 August 2026 the major vendors were checked in one pass, and the outcome differed sharply by vendor. Four served prices to an ordinary automated request. Four refused it. The table records what happened rather than smoothing it over.

A returned 403 is not an accusation. Bot filtering is routine and these companies publish the same documents to humans in browsers. It has one concrete consequence worth naming: an answer engine summarising AI pricing cannot read those pages either, which is precisely how a stale price survives for years while being repeated by everyone. That is the reason this site treats a blocked page as a gap to disclose rather than a cue to copy the figure from elsewhere.

Why an in-app purchase catalog is used as the fallback

An in-app purchase catalog is a first-party disclosure, not a secondary report. When a vendor's web pricing page refuses an automated request, the fallback is that same vendor's own App Store listing, where the tiers and prices it bills consumers are published as a legally required disclosure. It is a different surface with different limitations, and the pages say so: an app listing shows what a company bills through the store, not what it bills everywhere, so business seats and web-only promotional tiers never appear there. It also carries the platform's commission, which is why a plan can be priced higher inside an app than on the web, a gap our per-vendor pricing pages state explicitly where it exists.

When a source is known to be unstable

Three sources are recorded as unreliable to read once, and the affected figures carry the warning rather than a confident number.

• xAI lists the same tier name more than once at different prices. In the catalog read on 18 August 2026, three tier names appeared a second time at exactly ten times the first price. A single scrape therefore picks whichever entry happens to come last, so the per-tier rate limits and feature splits below the top tier stay marked as unpublished instead of guessed.
• Microsoft 365 Copilot scrapes as an oscillating figure, moving between a monthly rate and an annualised monthly equivalent as the annual tier appears and disappears from the page. The fact record notes the oscillation so a routine scrape never gets reported as a price change.
• Poe is recorded with no safe value at all. The existing corpus states one figure in 75 places and a slightly different one in 7, for the same tier, which is a genuine contradiction rather than two plans. Until it is read from the vendor, the record carries both variants and no value.

What gets omitted rather than estimated

A figure that cannot be read from a first-party source is left out, and the omission is stated where the figure would have been. This is the rule that costs the most traffic and it is the one most worth keeping, because a comparison page that quietly fills its gaps is indistinguishable from one that has no gaps. Four current examples:

• Our page on team and business plans omits OpenAI's seat price entirely, because openai.com, chatgpt.com and the OpenAI help center all refused the request and business seats do not appear in a consumer app catalog.
• Our cost per answer index runs its arithmetic with ChatGPT absent from the table, for the same reason, rather than importing a widely repeated figure it could not check.
• Our Perplexity pricing page states the free tier's daily cap as unverified rather than repeating a limit that has changed before.
• Our DeepSeek comparison publishes no DeepSeek token prices, because that vendor's API documentation refused connections from our network on the same day.

Why this site publishes almost no benchmark scores

A benchmark percentage that cannot be traced to a run someone else can repeat is a claim, not a measurement. Published model scores fail on three counts for a buying decision: they go stale within weeks because labs ship faster than any comparison can be re-run, they travel between articles detached from the methodology that produced them, and they measure a proxy for the thing a reader actually cares about. Our model comparison that publishes no benchmark score was reworked on that basis. The earlier version of it carried four tables of benchmark percentages with no traceable source. They were deleted rather than re-dated, and vendor rate cards read on a named day took their place.

Four kinds of number survive that filter: capabilities a vendor documents, prices a vendor publishes, limits a vendor states, and measurements this site runs and describes itself. The one composite ranking on the site names its inputs and its date rather than presenting a bare score, and the composite model leaderboard is dated April 2026, which is old enough that it should be read as a snapshot rather than a current verdict.

This rule is not yet applied across the whole corpus, and pretending otherwise would defeat the point of the page. About thirty older articles still carry benchmark percentages inherited from a generation template that hardcoded scores without attaching a source. Those pages are being reworked one at a time. Until a page has been, treat any bare percentage on it as unverified, and prefer the pages that name a source and a date.

How our own numbers are held to the same test

The strictest application of these rules is to figures about Perspective AI itself. The subscription price is settled and checkable: Perspective AI starts at $14.99/mo with no unpaid tier, and $49.99/mo is the tier above it. Both are recorded against the product's own documentation.

Three other figures are not settled, and the honest thing is to say which. The size of the model catalog and the monthly allowance included with the entry plan are both recorded as provisional, meaning one internal surface states a value while another disagrees or is silent. The number of agents each plan includes is recorded as unverified, with no public source at all. The standing instruction is to write around all three rather than pick the flattering variant, which is why no such number appears anywhere on this page, and why the site's structured data carries the prices but no counts. Pages that still state one predate the instruction and are being reworked alongside the benchmark pages above.

The same discipline applies to measurements this site produces itself. Where we publish our own numbers about model behaviour, the method sits next to the figures and the inputs are downloadable, so the refusal rate index and its published prompt set can be re-run by someone who does not work here. It also carries its own conflict of interest disclosure, because we sell the subscription that hosts the models it measures.

The automated checks and what each one blocks

Five checks encode these rules as code, and two of them fail the build. A rule that lives only in a document decays; a rule wired into the build either holds or breaks the deploy. The distinction between the two columns on the right is the honest part of this table.

Corrections policy

A wrong value is retired, not overwritten. When a figure changes or turns out to be wrong, the old surface form is moved into a permanent retired list before the new value is recorded with a fresh source date. The retired list is what the fact check greps for, so any page still carrying the old number fails the check rather than sitting quietly. That is the mechanism by which one correction propagates instead of being applied to the page someone happened to notice.

Four further rules govern what happens after a correction:

1. Dates move only when something moved. A page's last-updated date changes when its content changes, not on a schedule.
2. Dated measurements expire. Our own refusal measurements go out of the index and are marked non-indexable once they pass 45 days without a refresh, rather than being left to age quietly.
3. Removed pages redirect. Pages retired during the 2026 cleanup were given redirects to their surviving replacement, currently 376 rules, rather than being deleted into a dead link.
4. Corrections ship with the next deployment. This is a static site, so a fix lands when the site next builds. There is no live edit and no silent hotfix.

How to report an error

Email support@perspectiveai.xyz with three things: the page URL, the figure you believe is wrong, and the source you read. The third item is the one that makes a report actionable, because the correction has to be recorded against a source and a date like every other value. Reports about a competitor's pricing are as welcome as reports about ours, and a report that we have understated a competitor is worth more to this site than one that flatters it.

The same request applies to reasoning rather than figures. If a verdict on this site looks like it was reached by the commercial interest rather than the evidence, that is the most useful thing you can tell us, and the editorial standards page sets out the specific places where that interest bites.

What this methodology does not cover

Three limits, stated so nobody has to discover them.

• It covers facts, not verdicts. Prices, published capabilities and stated limits are checkable. Which tool is better for your work is a judgement, and the judgement carries the conflict of interest set out on the standards page.
• Not every record was re-read on its stated date. Several competitor prices in the store were established by agreement across the existing corpus rather than by a fresh vendor read, and those records say so in their own notes. They are re-verified as the pages carrying them are reworked.
• Articles published before July 2026 predate these rules. The rules were adopted on 24 July 2026 and apply to everything written since. Older pages are being brought up to them in order of traffic, not all at once, and a page that has not been reworked yet looks exactly like one that has until you check its sources.

Where these rules apply most visibly is the pricing work, since that is where vendors are least legible on purpose. The plan-price table every page here checks against carries the sourcing notes inline, and the pricing guides hub collects the per-vendor breakdowns built the same way.