What a link has to be to count, and what gets turned away no matter how well it scores. Nobody is asked to follow these on the honor system. The software checks.
Every link in the network has to survive being read by a person who does not know it was placed. That is the whole standard, and everything below follows from it. It is also the standard Google's own spam policies describe wanting: a genuine editorial citation, not a link traded as a favor.
It is a citation, inside a real article
The link sits in a body paragraph of a genuine piece of writing, supporting a claim that paragraph is already making. Not a resource list, not a links page, not a footer.
Content extraction locates the main article; links in headers, footers, sidebars, author bios, comments, and blogrolls fail.
The sentence would survive without the link
If the sentence exists only to carry a link, it is not a citation. Remove the link and the sentence should still be worth publishing.
Checked by a model against a fixed rubric on every submission.
The anchor is descriptive and reads naturally
Anchors are drawn from the target's own language: a title fragment, an entity, a descriptive phrase. They have to fit the sentence they sit in, in the tense and voice the author is already using.
Anchors are 1 to 9 words and screened against a commercial-term blocklist.
Dofollow, indexable, and findable
No rel="nofollow", "sponsored", or "ugc". The page returns 200, is not noindexed, and is reachable from the site's own navigation or sitemap.
Verified against both the served HTML and the rendered page, so links injected by JavaScript do not count.
The two articles are about related things
Not just the two paragraphs. A celebrity profile that happens to mention allergies is not a home for a supplements citation, however similar those two paragraphs look. Adjacent subjects are fine and expected. A piece on drones in construction citing one on site surveying is a good link. Unrelated subjects that share vocabulary are not.
Relevance is scored at both the passage and the whole-article level, plus a curated map of neighbouring topics.
These are refusals, not preferences. Most exist because the pattern is a recognisable footprint: the thing that makes a link network detectable, and therefore a liability to everyone in it.
Sales-word anchors
No "best", "top", "cheap", "buy", "hire", pricing terms, "near me", or "[service] in [city]" patterns. Ever, on any plan, for any price.
Blocklist plus classifier, re-checked on every monthly link check.
Linking back and forth
If you have linked to a site, it cannot link back to you for 365 days, and the reverse. Two sites trading links directly is both an obvious footprint and something search engines call out by name. Longer chains (you link to B, B links to C, C links to you) are fine and are what the matcher prefers.
Blocked at match time in both directions.
Repeat pairings
The same two sites cannot exchange again within 90 days. Two sites trading repeatedly is a pattern, not a relationship.
More than one link per article
One link to another subscriber per piece. Two if it runs over 1,800 words. An article carrying several placements is the single most recognisable link-scheme artefact there is.
A given article can also only ever earn credit once for a given domain.
Link-heavy articles
Roughly one external dofollow link per 300 words, counting all of them, not just ours. A piece carrying eight outbound commercial links fails even if only one came from here.
Thin content
Articles under 150 words cannot host or receive a citation. Collection pages need 250 words of real copy to count as anything other than a grid.
Placing links faster than you publish
Verified placements are capped against your site's actual publishing history. A site that has published twice a month for two years does not suddenly place ten links in a week.
Sites owned by the same account
Your own projects never match with each other. Tokens are shared across the domains on your account (an agency earning on one client can spend on another), but an account that never gives cannot receive.
Giving is judged across the whole account. See how the giving score works on the tokens page.
Programmatic and templated sites
Sites that generate pages from a template and a data source, like /plumber-in-{city}/ across four thousand cities, are refused. The signals are page growth far beyond what the site actually writes, and a large share of URLs collapsing into a handful of shapes. A single templated family over 200 pages is enough on its own.
Growth is measured against your own site, never as a page-per-day ceiling. A large publisher adding a thousand pages a day is publishing, and is not flagged. Multiplying your page count 3× in thirty days goes to a person; 10× with at least 1,000 pages added is refused automatically, and the refusal names the two numbers so a migration or an archive import can be appealed.
Catalogues, pagination, and utility pages
Product pages, page 2 and beyond of any listing, carts, checkouts, tag archives, and legal pages are never indexed and never eligible. A store is welcome; its catalogue is not the part that participates.
On by default from the day a domain is approved, and one switch on the domain's settings page turns it off. Unattended insertion is a separate opt-in that stays off unless you choose it. Worth understanding before you rely on either. Instead of citing something in a piece you are writing, we find a passage in an article you published months ago that could honestly support a citation.
On by default, and one switch turns it off
Adding links to older posts is a pattern search engines associate with link schemes, so every suggestion waits for you and nothing is ever changed on your site by us. If you would rather not see them at all, turn Existing Opportunities off on that domain's settings page.
Articles at least 7 days old
Younger than that and the edit is indistinguishable from writing new content, which is what the other path is for. There is no upper limit: the pages worth citing from are often the oldest ones you have.
10 open at a time, 1 per article, ever
You see up to 10 suggestions per domain, topped back up as you accept or dismiss them. One you leave alone expires after 14 days and another takes its place. No article ever accumulates citations, however many you work through.
You approve every one
Proposals wait for your click, showing the exact paragraph, anchor, and sentence. Unattended insertion is a second, separate opt-in that stays off unless you choose it.
Always reversible
A revision is saved before any edit, so every insertion can be undone from your post history. If the paragraph changed since the proposal was made, the edit is refused rather than applied to the wrong place.
Same standard as everything else
The paragraph still has to support the claim, the anchor still has to be descriptive, and the link still goes through independent verification before any tokens are granted.
Next: What we check →