AI Features Policy

1.What the Copilot is

The Copilot is a convenience tool that answers questions about software using our own published review data. It is not a reviewer, not a salesperson, and not an authority on anything.

It exists so you can ask "Notion or Obsidian for a solo founder" in plain language and get an answer grounded in the same database that powers our comparison pages.

2.How it works

The Copilot runs in one of two modes, and it always tries the first one.

• Local retrieval first. We search our own review database for tools that match your question. If the match is confident, it answers entirely from our published data and no external model is contacted.
• Model synthesis second. Only when local retrieval is not confident enough does it send your question to OpenRouter, a United States company, which routes the request to whichever language model it selects. The model is instructed to quote prices and features only from the context we supply, never to invent a price or a plan, never to name competing review sites, and to keep answers short.

Outbound calls are capped at 700 completion tokens, and the context we send is capped at roughly 4,000 tokens. Those caps are cost and safety controls, not quality controls.

In practice this means most questions are answered without your text ever leaving our server. The moment retrieval is not good enough, it does leave.

3.What leaves your browser, and where it goes

When the Copilot falls through to model synthesis, your question text is sent to OpenRouter over HTTPS and processed in the United States, and potentially onward to whichever model provider OpenRouter routes to. We do not control which provider handles it.

This is why AI use has its own consent category in our consent panel, separate from analytics and affiliate cookies. It is a transfer of your content to a third party in another country, not a cookie, and you should be able to agree to it knowingly.

• Do not type anything confidential. Client names, unreleased financial figures, roadmap details, credentials, or anything covered by a non-disclosure agreement. Assume anything you type may be read by a service outside the European Economic Area and the United Kingdom.
• Turn it off and nothing breaks. If you decline the AI category, the Copilot still answers, but only from our own published data, with no external model involved. You lose nothing except the ability to ask a question retrieval cannot answer on its own.

4.What we keep, and what we do not

• Your questions are not written to our servers. They exist in memory only while the answer is being generated, and are discarded when the request ends.
• A copy lives in your own browser. We store your conversation in local storage, under the keys `gopickstack:ai-chat` and `gopickstack:ai-chat-sessions`, so you can see your own history. We keep at most 20 saved sessions. That data is on your device, we never receive it, and clearing your browser data removes it permanently.
• Questions are not joined to your email. We cannot tell who asked a question unless you separately subscribed to the newsletter, in which case the two datasets are not joined.
• We do not review your questions to improve the product. They are not used to build a profile of you.

5.What it can get wrong

The single most important thing to understand about this feature: it can be wrong, and it can be wrong confidently. A fluent, well-formatted answer is not evidence that the answer is correct.

• It can misstate a price. Verify anything financial against the vendor's own pricing page, not the Copilot.
• It can attribute a feature to the wrong plan or the wrong tier. Feature flags change constantly and the model is working from context we published on a 15-day cycle.
• It can produce a plausible answer with nothing behind it. Where retrieval is weak, the model may fill the gap rather than say it does not know. This is the main failure mode and it is why we cap how much the model is allowed to do.
• It does not independently test software. It reports what we published. If we did not test something, the Copilot cannot tell you about it.
• It is not professional advice. Not legal, financial, tax, accounting, or procurement advice. A consultant who quotes our Copilot as the basis of a recommendation is misusing it.

Every Copilot answer surfaces the source reviews it drew on. If you are about to make a purchase decision, open the cited review and read the section that matters. That is the whole point of citing it.

6.Our content, and yours

The Copilot does not have opinions we do not have. It is constrained to the scores, pricing, and verdicts we have already published, and where it goes beyond them it is extrapolating rather than reporting.

The system prompt instructs it never to name a competing review site, and never to invent a price or plan. Those are hard constraints we set deliberately, because a review site that recommends a rival is a review site nobody can trust.

7.Acceptable use

You agree not to use the Copilot to:

• Break the law, or produce content that does.
• Impersonate a person or organisation, or generate fraudulent material.
• Harass, stalk, defame, or threaten anyone.
• Distribute malware, or attempt to attack the service or any system it connects to.
• Scrape, enumerate, or overload the endpoint, or attempt to circumvent the rate limits or the daily ceiling.
• Submit personal data about someone else without their consent.

We may refuse or throttle requests that breach this section. Deliberate rate-limit evasion is treated as a breach of our Terms of Service, not merely an API preference.

8.Availability and cost limits

The Copilot is offered without any uptime guarantee. It is a convenience, and it can be disabled or degraded at any time without notice.

We apply a hard daily ceiling on the number of upstream model calls the site may make, and a cap on how many can run at once. When the ceiling is reached, the Copilot does not fail: it stops calling external models and answers from our own published data, and it tells you it is busy.

A note on rate limiting, stated honestly: the site is served directly from a host with no reverse proxy in front of it, which means IP-based limiting is best-effort only and a determined caller could work around it. The daily ceiling is what actually bounds our cost, which is why it is the control that matters.

9.Changes to the technology

The model provider, the routing, and the context we supply all change over time. When a change is material, meaning it affects what the Copilot can know, how it is routed, or where your text is processed, we record it in our public changelog and update this page.

If we ever change who receives your questions, that is a material change and we will say so here before it takes effect, not only in a silently updated subprocessor table.

10.How this fits with our other terms

This policy is part of our Terms of Service and should be read with our Privacy Policy, which explains the lawful basis for processing your questions and lists every third party that touches your data in our Subprocessors register.

It is governed by the laws of England and Wales. Your statutory rights under the law of your own country are unaffected.