AI Transparency

How a recommendation is made

Ask FetchIt runs the same pipeline every time: it turns your words into a structured requirement (category, budget, priorities, usage, how long you’ll keep it), filters FetchIt’s catalog by your hard limits, scores every eligible product for you, checks risks and evidence strength, then picks up to three options — Best Match, Best Value and Best Alternative.

This catalog answer runs on deterministic, rule-based logic over FetchIt’s verified product dataset — the same input always gives the same result, and FetchScore, FetchMatch, FetchRisk and FetchConfidence come only from it.

You can always see what FetchIt understood (“What FetchIt understands”, including the phrase each item came from), correct it, and watch the recommendation recalculate.

Live AI research

Alongside the catalog answer, Ask FetchIt runs live research: an AI model (Anthropic’s Claude) searches the web at the moment you ask — Indian retailers, brand stores and reviews — and recommends specific products with their trade-offs. It can answer about products that aren’t in FetchIt’s catalog, and it is given the catalog so it can point you to products FetchIt has verified.

Every link it shows was returned by its web search during your request, and a price is only shown with the page it was seen on (“seen on flipkart.com”). FetchIt removes any link, price or product the search didn’t support before you see it. Live results are still AI-generated from web pages that can be wrong or out of date: check the price and specs on the retailer’s page before buying. Live picks don’t get FetchScores — only catalog products do.

Live research runs when a signed-in member chooses it and uses FetchCredits; the cost is shown before anything is used, and a request that fails is never charged. There are also per-network and daily limits to control cost; when one is reached you get the catalog answer only. Identical questions within 24 hours are answered from a saved copy at no charge.

FetchIt uses Anthropic’s Claude models. Simpler tasks (like follow-up questions) use a smaller, faster model; when overall usage is very high, a request may be handled by a smaller model rather than refused.

Follow-ups, comparisons and your research history

Your live research is saved to your research history so you can come back to it. Follow-up questions are answered only from that saved research — no new web search — and when a question needs information the research doesn’t have, the answer says so rather than guessing.

Multi-product research compares products you name, live. The same rules apply: every price and link must come from that search, and a product, price or “winner” the search didn’t support is removed before you see it.

How feedback improves answers

When you rate an answer, FetchIt’s editors see it alongside others in the same category. They write short research guidance for that category — for example, “compare plain selling prices, not bank-offer prices” — which the live research then follows. An AI may suggest guidance from the feedback, but a person reviews and approves every change. Guidance is about how to research, never product facts or brand preferences. Your feedback isn’t used to train AI models.

AI-researched catalog entries

Some catalog products are first researched by AI: specs from the brand’s page and current prices from named stores, with every link coming from the research’s own search results. An editor checks and approves each entry before it’s published. Their FetchScores are spec-based estimates and are labelled that way, the same as other researched entries.

FetchScore vs FetchMatch

FetchScore answers “how objectively good is this product?” It’s a weighted score out of 100 using category-specific dimensions — camera and battery for phones, cooling and efficiency for ACs — and is the same for everyone.

FetchMatch answers “how well does it fit you?” It re-weights the same dimension scores by your priorities (Low, Medium, High, Critical), then accounts for budget fit, expected ownership period, usage and sizing (for example, an AC rated for a smaller room than yours is marked down). It changes as soon as your priorities change.

FetchRisk and FetchConfidence

FetchRisk lists reasons a product might disappoint: price volatility, product age, software support, reliability, service network and recurring owner complaints. Each risk cites the data it came from. Every product page also has a mandatory “Reasons you may NOT want this” section.

FetchConfidence reflects how strong the evidence is: number of owner ratings, expert sources, retailer coverage, completeness of the spec sheet, how recently prices were checked, and how close the top options are. Data that hasn’t been verified against live sources — including FetchIt’s current demo dataset — is always rated Low confidence, however much of it there is.

Clarifying questions and challenges

FetchIt asks a clarifying question only when the possible answers would lead to a different Best Match, and never more than three per session. You can always skip and see recommendations.

“Challenge FetchIt” answers are calculated by re-running the scoring under the changed assumption (a bigger budget, a different top priority), not written as opinions.

What FetchIt will not do

Invent specifications, prices, discounts, review statistics, retailer availability or sentiment; present unverified or outdated data as confirmed; hide a product’s known weaknesses; or claim to know future prices. When FetchIt lacks enough verified information, it says so instead of recommending.

Commercial independence

Retailer relationships and affiliate commissions are not inputs to FetchScore, FetchMatch, FetchRisk or FetchConfidence. FetchIt does not currently earn commission on retailer links; if that changes, it will be disclosed on the Affiliate Disclosure page and next to the links themselves, and it will still not affect any score.

Your data

Your Ask FetchIt requirements are kept only in your browser tab’s session storage and are cleared when you close the tab. Your watchlist, saved decisions and recommendation feedback are stored in this browser only.

For live research, your question and FetchIt’s reading of it (category, budget, priorities) are sent to Anthropic to generate the answer. Don’t include personal details in questions. FetchIt’s server keeps the answer for 24 hours under a code derived from the question (not the question text itself), and counts searches per visitor for one hour using a scrambled form of your IP address that can’t be read back.

FetchProfile (your saved preferences) exists only if you create it, either at My FetchProfile or by ticking items under “Remember these preferences” — nothing is pre-selected. When it fills a gap in a question, Ask FetchIt labels each item “from your FetchProfile” and lets you ignore it for that question. Anything you say in a conversation overrides it. You can pause or delete it at any time; deleting removes every field.

Limits

Match percentages and verdicts are estimates from available data, not guarantees of how a product will perform for you. Always confirm the final price, warranty and return terms on the retailer’s site.