How PickDuel works

You enter a website. We read up to three of its public pages (usually the homepage, pricing and product pages), run one web search to find who it competes with, and do the same for the closest rival.

A language model reads that evidence and rates specific signals from 0 to 5. The final score is calculated by fixed code, not by the model, using the weights below. The written comparison is generated from the same evidence and links back to its sources.

The score reflects how clearly each business communicates on its public website. It is not a judgement of product quality, and we don't claim to have used either product.

Community backing is separate. Backers never change the PickDuel score, and we never add votes ourselves.

We compare what a company promises on its own website with what public reviews and discussion keep saying about it. The promise is quoted word for word from the company's site, and the quote is checked against the page before anything is published.

The evidence is the company's most recent Trustpilot reviews, found by its own domain, plus public discussion from forums and developer communities that express an opinion about the product. We never show reviewers' names or copy their reviews: each point links to the original.

A language model only sorts that evidence by topic (for example pricing, support or ease of use) and marks whether each item praises or criticises that topic. Fixed code then does the counting. A promise is strongly reflected when at least 80% of related mentions are positive, mostly reflected from 65%, mixed from 40% and weakly reflected below that. We need at least 5 related mentions to judge a promise, 3 to show a topic and 20 overall to publish a check at all. Otherwise we say the evidence is insufficient.

Review sites over-represent people with strong experiences, especially billing and support problems. A check shows what reviewers say, not what every customer thinks. Every check shows the date it was made.

When someone checks a company we don't have yet, this runs automatically in the background and appears on the page within a few minutes. Automation changes who starts the work, not the standard: the same thresholds decide whether anything is published at all, and if the public evidence is too thin we publish nothing and say so. Our operators review what goes out and can remove any check.

Contractors are checked on what a homeowner looks for, not on a software rubric, and never marked down for not publishing a price list. Each check is a fixed rule over the pages we read, and every check that passes shows the exact sentence it matched.

• 25Service clarityServices named clearly (15) and emergency or same-day service (10).
• 25Trust proofLicensed and insured (8), a real manufacturer or trade certification (7), years in business (5), customer reviews on the site (5).
• 15Service areaWhere they work, stated on the site.
• 20Contact and next stepA phone number or call action (10) and a free estimate or inspection (10).
• 15Warranty and financingA warranty or guarantee (8) and financing or payment plans (7).

Businesses come from public Google Maps listings. What customers praise and question comes from their most recent Google reviews, grouped by topic with the same fixed counting rules as above. Google reviews tend to skew positive. Local arena pages count only businesses we actually analysed (“among N businesses analysed by PickDuel”) and are sorted by review count, never presented as a ranking. None of this measures the quality of anyone's work.

Hosting is sold at a first-year price and billed at a renewal price for every year after. The gap is often large — we have read plans that advertise $2.99 and renew at $17.99 — and it is the number that decides what the hosting actually costs you, so we put it first.

Plans come from the host's own pricing page. A price is stored only if it is printed on that page: if a renewal price is not published we say so and leave it blank, rather than assuming the industry norm. Where a host answers our reader with a block page, we say we could not read it and publish nothing else about its prices.

“First two years” is arithmetic, not a forecast: the advertised price for the stated term, then the renewal price for the rest of the twenty-four months. The list on our hosting page is ordered by the size of the increase. That is not a ranking of how good the hosting is, and nothing on it can be bought.

We never estimate a price. Everything in the pricing block is read from the company's own pages: how it bills, the lowest price it actually prints, whether there is a free plan that doesn't expire, any stated minimum commitment, and the charges it says can increase a bill.

Plenty of businesses have no single price. A cloud platform or a telco can publish thousands of rates, and a “contact us” company publishes none. Inventing an average for either would be a lie, so we say there is no single price, name what the meter charges against, and list the extras that catch people out. Next to that we put what reviewers repeatedly say about the bill, counted by the same rules as every other theme.

A lower starting price is not a verdict. We say which side prints a lower number and leave the judgement to you, because what a business actually pays depends on its size and usage, which we don't know.

A shortlist is the set of businesses a buyer would realistically weigh up together. It comes from what already exists: comparisons on PickDuel, the companies the public lists next to each other in search results, the market a company belongs to, and, for contractors, the businesses working in the same city. A name only appears if its domain was in the evidence we read, so nothing is conjured from a model's memory.

Every card has to carry a checkable reason: publishes a price, has a free plan, starts at a stated amount, is praised by a stated number of reviewers. A business with nothing factual to show against the company on the page is left off. The cards are not ranked by quality and no business can pay to appear on, or disappear from, a list.

The rules run in both directions. Any company on your shortlist also has you on theirs, for the same reasons. If a card is wrong about your business, use “report incorrect information” on your company page.

No account is needed. To count one backing per person we keep a random ID in a cookie, and we store network addresses only as one-way hashes.

We use Google Analytics to see which pages help people, with advertising features and Google signals turned off. Visitors in the EEA, UK and Switzerland are asked first, and nothing is stored until they accept.

Change analytics cookie choice