The links report is a sample, not a total
Google's links report is a sample of the links it already knows about, split across two tabs with two truncation limits. What that does to a link audit.
Google has never handed you a list of the links pointing at your site. It hands you a sample of the links Google already knows about, and this week it confirmed again that "not all links" includes not all spammy ones. So the honest output of a link audit is a pattern you can point at by name, never a percentage of a profile nobody can see.
What the report actually is
The Links report reads like an inventory. It is built out of two different samples and one export ceiling, and Google's own help documentation says so in plain language.
| Tab or table | What Google documents | What it is not |
|---|---|---|
| Latest links | The most recently discovered links to your site, up to 100,000 rows, sorted by discovery date | A ranking of your most important links. It is sorted by when Google found them |
| More sample links | "A sample of links to your site (up to 100,000 rows), extracted from the full set of links known to Google" | The full set. The tab name says sample and the definition says sample |
| Top linking sites, top linked pages, top anchor text | Aggregate tables. Single-table pages export up to 1,000 rows | A total. Above 1,000 rows the export is a truncation |
| Links you disavowed | They can still appear in the report | Evidence that Google is still counting them |
Read the second row twice. Google separates the links it knows from the sample it will show you, and the sample is what you can download. That distinction is the whole article.
It also means the two tabs are not interchangeable. Latest links is a recency window: what you see first is what Google found most recently, not what matters most. More sample links is a slice of unknown basis. Run this quarter's audit from one tab and next quarter's from the other, and you will be comparing two datasets that came out of the same tool, about the same site, and share no denominator.
The aggregate tables have their own trap. Top linking sites and top linked pages are single-table pages, which export up to 1,000 rows. On a site with more linking domains than that, the export is a truncation. Anyone who has sorted that file by link count and announced the largest linking domain has been reading a list that stops at a number Google picked.
Three questions people bring to the report
Every link question I get asked resolves to one of these three. The report is genuinely useful for one of them.
| The question | What the answer actually requires | Does the report answer it? |
|---|---|---|
| Did someone build links at us? | A pattern across linking domains, anchors and dates, at a level you can name | Partly. It shows presence, never volume |
| Are our links doing anything? | Referral sessions, branded search, movement on the linked URL | No. The report has no traffic, no effect and no crawl behaviour in it |
| Should we disavow? | A manual action, or a paid-link pattern you can evidence | Rarely. It is not the deciding input |
Only the first is a reporting question. The second belongs in analytics. The third belongs next to Google's own guidance, which has been consistent for years and was restated again this week by Search Engine Roundtable: Google does not show all links, does not show all spammy links, and has repeatedly said the disavow tool is not something most sites should be using. Google's position is that its algorithms discount link spam on their own, and that a disavow file applied too broadly can hurt more than the links it was meant to remove.
There is a second detail in that writeup that most audits miss, and it is the one that should end the argument: the links report shows links you have already disavowed. A row in the report is a record of discovery, not proof of counting.
The false denominator
The failure I see most often in client link audits is not a tooling problem. Someone exports a tab, groups by domain, counts the rows they do not like, divides by the row total, and writes that a third of the backlink profile is toxic.
Nothing in that arithmetic is wrong. Everything in it is wrong about the denominator. The row total is a sample size, not a profile size. It moves when you switch tabs, when you change the date range, and when Google discovers more links. So last quarter's percentage and this quarter's percentage came from different populations and cannot be compared, which means the trend line in the slide is decoration.
What makes this worse than an ordinary measurement error is how well it travels. It looks precise. It fits the shape of a deck. It produces a clean action item. And nobody upstream can check it, because checking it means reading the help page for the tab that was exported.
My rule after enough of these audits: a number you cannot defend line by line does not go in a report. Toxic-link share fails that test by construction, not by degree.
What the sample can still prove
None of this makes the report useless. It makes it a pattern instrument rather than a balance sheet, and patterns are what you actually act on.
Three things survive the sampling problem. The names: a cluster of domains sharing a registrar, a naming convention or a template makes a network visible even when the row count is unknown. The anchors: an anchor distribution concentrated on one commercial phrase, arriving in a narrow window, is a pattern regardless of what fraction of the file it represents. The company it keeps: links that appear alongside the same set of sites across several properties tell you something about how they were acquired.
What does not survive is any claim about how much. Four hundred bad links in one export and four hundred in another are not the same finding, because the exports may cover different shares of the profile. Compare shapes, not counts.
The same mistake, now in AI-search reporting
This is not a links problem. It is what happens whenever a platform shows you a cut-down surface and files it under a report.
Search Console's AI surfaces are the current example. Every link inside an AI Overview reports the position of the block, not a position within the answer, so averaging those positions produces a number wearing a rank's clothes. We went through the mechanics in Search Console gives every AI Overview link the same position — the counting rule is documented, the AI Mode rule is not, and the impression history contains at least one logging error that was never corrected backwards.
A link inside an AI answer is also a different asset from a link in a ranked list, which changes how you read what AI Overview links are actually sending. Both cases land on the same discipline: a truncated surface tells you which way the current runs. It does not tell you how strong the current is. Use it to find patterns. Go somewhere with a counted population when you need a rate.
The four-step check before a link report leaves the building
- Write the question as a decision. "Do we disavow in Q4" has an answer. "How healthy is our profile" has a slide.
- Name the population and the tab. One line in the document: which tab, which date, what the row total represents. If you cannot write that line, you are not ready to write the percentage.
- Test the denominator before you divide. If the number is capped — 100,000 rows, 1,000 rows — say so next to the percentage or delete the percentage.
- Name the action. If no decision changes because of the number, you are producing a deliverable, not a measurement.
This is the part of an SEO and GEO engagement that has to be boring. Same tab, stated limits, exports kept, and no percentage that cannot be re-derived from the file that produced it. When Google tightens link spam, as it signalled it would in the September 2026 spam update, the visible change is a set of pages losing visibility. Not a number moving in your spreadsheet.
The rule
A sample can show you a pattern and cannot give you a rate. Google put that in the name of the second tab and repeated it this week. Two tabs, two population rules, two ceilings — and one question that survives all of it: what did I actually see, and what would I do differently because of it?
Frequently Asked Questions
Does Search Console show every link to my site?
No. The Links report shows links Google already knows about, and one of its two tabs is explicitly described as a sample extracted from the full set of links known to Google. Links Google has not crawled, and links on pages it chooses not to index, never reach the report at all.
Why does the links report show links I already disavowed?
Because the report is a record of links discovered, not a record of links counted. A disavowed link can still appear in the list. Seeing it there does not mean Google is still using it, and deleting the row is not a way to disavow it.
Can I calculate the percentage of toxic links from the export?
Not honestly. The row total is a sample size, not a profile size, so any share you compute from it is a share of an unknown population. The report can show you a pattern in the names, anchors and timing of links. It cannot tell you what fraction of your profile is bad.
