How GooDu orders what it lists

GooDu is an automated index of publicly reachable storefronts. Every listed entry is described with the same fields, and the order is produced by a fixed formula that can be re-run and re-checked. This page states why the order exists, how it is produced, which inputs it is built from, which inputs are excluded, and how manual action is bounded.

This page is a statutory disclosure, and it is deliberately organised around the three things that have to be disclosed. Under 《互联网信息服务算法推荐管理规定》第七条 the rules of an algorithm recommendation service must be made public; under 第十六条 the service must tell users that it applies such an algorithm and must publicly set out its basic principle, its purpose and intent, and its main operating mechanism — sections 1, 2 and 3 below, in that order. Section 6 states what is deliberately not published, and why.

1. What this order is for

The purpose is narrow, and it is worth stating plainly because a ranking is easy to mistake for an opinion:

• To put the entry that answers the query first. A visitor searching for a term should meet, on the first screen, the entries whose observed facts relate to that term — not the entries that have been around longest, or that someone would prefer they see.
• To make that choice explicable and repeatable. Every position is the output of the same formula over the same observed inputs, so a position can be reproduced and challenged. A position that cannot be explained is one that can be sold, and this one cannot be sold.
• Nothing else. The order is not an assessment of quality, not a rating, not a recommendation, and not a safety judgement. It says which entry is a better answer to the query, and nothing about the site behind it.

2. How the order is produced

Four stages, in order. The first three are the same for every entry; the fourth is the formula.

1. Candidates. Addresses are collected from public sources — search engines, aggregating directories and public listings. Nothing about any visitor takes part in this.
2. Probes. Each candidate is requested over HTTP by our own servers, and the response is recorded. A site enters the index only if a probe has observed it reachable at least once, and each entry is re-checked over time.
3. Fields. For an entry that passed, the listed fields are read from the site's own public pages and from domain-registration records. Every entry carries the same fields, computed the same way — no entry has a field that another entry does not have.
4. Order. The formula runs over those fields, then adds a bounded, non-personal variation (below). The result is the list, in order.

The bounded variation, and why it exists

On the public search pages, an entry's position carries a variation of at most ±15%, which in practice moves entries between neighbouring positions rather than across the list. Three properties matter:

• It is non-personal. The seed is drawn fresh on each request and mixed with the search term — never with the visitor. There is no profile, no IP-based or user-agent-based adjustment, and no memory of previous searches, so the order is not a function of who is asking.
• It is re-rolled per request, and stable within one response. One response carries one seed: paging through it, or changing how many rows per page, does not reshuffle what you are looking at. Running the same search again does reshuffle it — that is the point (see the third property).
• Its purpose is to prevent a permanently fixed front position. An index whose first position never changes is one where that position becomes worth acquiring, and where the ranking rule stops being re-examined. The variation makes that position temporary by construction, while leaving clearly more relevant entries ahead of clearly less relevant ones.

The variation applies only to the public search pages. Internal evaluation runs and the command-line tools sort with no variation, so that their output is byte-for-byte reproducible — a measurement that changes between runs cannot be used to judge anything.

Is the order the same for everyone?

It is not personalised, which is a different thing from being identical. The order does not depend on who the visitor is, where they are, what they searched before, or what device they use: no profile is built, no IP-based or user-agent-based adjustment is applied, and the interface language does not change the order. But the bounded variation described above is drawn again on every request, so two people searching the same term will generally see two different orders — and neither of those orders is a function of who they are. This service does not target results at personal characteristics, so there is no personalisation to switch off — see Privacy.

3. What the order is built from

Only observed, re-checkable facts enter the formula:

• Whether our own probe reached the site, and when it last did. A site is listed only if a probe has observed it reachable at least once.
• When GooDu first recorded the site. This is our own record of first observation; it is not the age of the site, and it is not the domain's registration date.
• Facts read from the site's own pages — for example whether a page states automatic delivery, which payment channels were observed, and which support channels were observed.
• Domain registration facts from RDAP/WHOIS, when they are available.

None of these is a judgement about quality. GooDu does not compute and does not publish an assessment of, rating of, or recommendation for any indexed site.

4. What never affects the order

• Payment. There is no paid placement. No advertiser, no sponsor, and no paying site can move an entry. Any future paid placement would have to be labelled, marked as rel="sponsored", and physically separated from organic results.
• Manual promotion. Manual action can only remove an entry. There is no action that raises one.
• What a site says about itself. Names, headings, slogans and promotional claims written by a site are not inputs to ordering, and are not carried on listing pages.
• Our own interest. Sites operated by us, affiliated with us, or partnered with us are not given any position that another site would not receive under the same observed facts.
• The visitor. No personal characteristic is an input — see the end of section 2.

5. Manual action, and how it is bounded

There are exactly two grounds on which a human changes what is listed:

1. Lawful removal. A rights holder submits a notice that meets the requirements in 《民法典》第一千一百九十五条, or a competent authority issues a legally binding order.
2. Removal required by law. Content that applicable law requires us to remove.

Every manual action is logged — who, when, which domain, and why — and that log is available to internal audit. There is no manual action that improves a position.

6. What we publish, and what we do not

• Published: the purpose, the mechanism, the inputs, the exclusions, and the bounded variation — this page.
• Not published: the numeric weights. A published weighting turns an index into a scoring guide that can be gamed; sites that do not game it would then be pushed down by sites that do. Keeping the weights internal keeps the formula from becoming an attack surface. The disclosure obligation is one of principle, intent and mechanism — a disclosure of weights would defeat the purpose it serves.

For the same reason this page is not treated as marketing material: it is written so that a position can be checked against it. If a position cannot be reconciled with what is written here, what is wrong is the position.

《互联网信息服务算法推荐管理规定》第十五条 also prohibits using a recommendation algorithm to place unreasonable restrictions on other information service providers, or to obstruct or disrupt their lawful services, so as to engage in monopoly or unfair competition. The exclusions above are how this service stays inside that line.

7. Algorithm filing and how to reach us

Under 《互联网信息服务算法推荐管理规定》第二十四条, a service with a public opinion or social mobilisation function must file with the competent authority within ten working days of starting to provide the service, and under 第二十六条 the filing number must then be shown in a prominent place on the service. The category this service falls under is retrieval and filtering (检索过滤类, defined in 第二条 of the same rules).

This deployment has not yet been publicly launched, and the filing is therefore not yet made; this page is published in advance of launch so that the disclosure exists before the service is reachable by the public. The filing number and a link to the published filing information will be placed in the footer of this site, and stated here, as soon as the filing is made.

Rights notices, takedown requests and reports are handled on a separate page — see Rights notices & reporting.