RankBrain backlinks: what the claims get wrong
Half the advice about machine-learning ranking is repeated without a source. Here is each popular claim next to what Google actually said, and whether link data can check it.
Verdict
RankBrain is a query system, so link tricks aimed at it miss
Google described RankBrain in October 2015 as a way to handle searches it had never seen. Links were not part of that description. Anyone selling a RankBrain link package is selling a story.
- Documented: RankBrain helps match unusual queries to pages.
- Documented: BERT, added in October 2019, reads word context in queries.
- Not documented: any link formula inside either system.
Myth against reality
| Claim you will hear | What is documented | Can link data test it? |
|---|---|---|
| Build links with the exact anchor so RankBrain learns your page | Nothing says RankBrain reads anchors; it interprets queries | Partly: see which anchors your rivals have, then judge |
| BERT punishes keyword-stuffed anchors | BERT is a language model for queries and passages, not a penalty | No; Penguin and SpamBrain are the link-side systems |
| More referring domains always beat better content | Google lists many signals and never ranks links first | Yes, on your own sample: compare counts to positions |
| Helpful content needs links to count as helpful | The system, from August 2022, scored content; links were not its input | Only as a side check on who cites you |
| A low DR means ML systems distrust you | DR is a third-party style metric; Google does not use it | No; ours is a log scale of referring domains |
Dates worth knowing
October 2015: Google confirms RankBrain. Then 2018 brings the August core update, and October 2019 brings BERT. August 2022 gives us the helpful content system, followed by SpamBrain's link spam update that December.
In March 2024 helpful content became part of core ranking and Google ran a spam update the same day, 5 March. Reading those entries side by side shows that the link-side and content-side systems move on different clocks.
Keep that timeline beside any traffic chart. Most claims of ML causation vanish when the dates do not line up.
Where the myths come from
Machine learning sounds like magic. People fill the gap with whatever they already sell. Link vendors say links, content agencies say content.
Google's own wording is narrower. In March 2016 a Google search lead named links, content and RankBrain as the three top signals, yet the company never published how RankBrain is computed or what it weighs against links.
So read the careful version: a system exists, it matters, and its inner workings are unpublished. Ask any claim to name its test. If it cannot, move on.
What a marketer can actually do with this
Skip the theory and run a comparison. Our worked example shows the calls, using invented sites so nobody mistakes it for a finding.
Then look at the other side. If a drop followed a wave of strange anchors, the spam runbook tells you whether to act or wait. Often you wait.
BERT shows over-reading well. Hype ran high. It reached English searches in the US in October 2019 and over 70 languages by December, and many sites saw nothing at all. Rewriting every page for it was a waste of a quarter.
Pick 30 queries you track. Pull referring-domain counts for the top three results with compare, which is one credit per call. If counts bear no relation to rank across the 30, your theory about links and ML is already in trouble.
How to read a link-and-ML claim
- Find the date and the source; Google announcements carry both.
- Check whether it speaks of queries, content or links.
- Look for a sample larger than one site.
- Note whether the author sells what the claim recommends.
- Repeat your own pull after the next monthly index refresh.