Since AI Overviews launched, nearly every talk I give ends with someone asking whether backlinks still matter. My answer: yes, though less than the link-building industry would like you to think, and the work around them has changed more in the last two years than in the decade before.
Google still uses links, and has spent years teaching its systems to ignore manufactured ones. AI search adds a second question: you used to ask who links to you, and now you also ask who talks about you, and whether AI answers notice.
This is the hub of a five-part series on an AI-powered backlink strategy: what Google and the research say about links, what AI search changes, and a workflow that uses AI where it helps. The running example is my own site. On 7 November 2026 I connected Claude to the DataForSEO MCP server and pointed it at mlforseo.com, and it was as messy as real profiles are.
November 2026: a 2021 guide, rewritten. This post replaces the backlink strategy guide on my personal site. I’ve kept what still holds, corrected what didn’t age well (follow links as “critical”, Domain Authority thresholds, HARO, the fear of toxic links) and added the research and the AI workflow.
On this page
The series
Four companion posts do the hands-on work:
- How to do an AI-powered backlink audit: the signals, tools and research for judging link quality, and how far to trust each third-party metric.
- Backlink analysis with Claude and the DataForSEO MCP server: a no-code tutorial with the exact prompts I ran on mlforseo.com.
- Score backlink quality in Python: a notebook that pulls DataForSEO data, adds embeddings for topical relevance and asks Claude to label each link with a reason.
- After the backlink audit: what to ignore, when (rarely) to disavow, how to reclaim broken links and which prospects to pursue.
Do backlinks still matter in 2026?
What Google says
Google’s link best practices say it uses links as a signal for the relevancy of pages and to find new pages to crawl, and its ranking systems guide still lists link analysis systems and PageRank among the systems in use.
It has also been turning the volume down. In March 2024 its documentation stopped calling links “an important factor” and called them “a factor”, as Search Engine Journal reported. Gary Illyes has said links are not even in the top three ranking factors, and that Google needs very few links to rank pages (a line he later said he shouldn’t have used). My read: a handful of relevant links from real sites does more than hundreds of weak ones.
What the papers and patents describe
The research explains why some links count more. A caveat: a patent describes an idea a company protected, which doesn’t prove Google uses it.
- PageRank (1998). Brin and Page’s original paper modelled a random surfer clicking links. PageRank is calculated per page, and anchor text is credited to the page the link points to.
- The reasonable surfer. This expired Google patent describes a surfer who follows some links more often than others, based on features such as the link’s position (body, sidebar, footer), font size, the anchor words and how well the linking page’s topic matches them. So a link in a relevant article beats a footer link.
- Distance from trusted seeds. Another Google patent ranks pages by how few link hops separate them from a diverse set of trusted seed pages.
- TrustRank and Hilltop. TrustRank (Stanford and Yahoo!, 2004) spreads trust from a small, human-reviewed seed set, because good pages seldom link to bad ones. Hilltop weights expert pages that link to many independent sites on one topic. Neither is confirmed as part of Google.
- Link spam detection. By 2008 researchers had classifiers using link features alone that performed comparably to content-based ones, and a Google patent on historical data describes a sudden spike in backlinks as a possible spam signal.
The 2024 Content Warehouse leak shows fields that echo these ideas (a quality tier for an anchor’s source page, anchor spam demotion), though, as Mike King notes in his analysis, nobody knows their weights. Together, the research points where common sense does: links from relevant, independent, trusted pages, placed in the body of real content, count for more.
How Google treats link spam now
Google’s spam policies define link spam as links created mainly to manipulate rankings, from bought links and link exchanges to optimised anchors in guest posts and press releases. Penguin joined the core algorithm in 2016, devaluing spam instead of demoting whole sites, and Google now lists it as retired. Since the December 2022 link spam update, SpamBrain, Google’s AI-based spam system, neutralises unnatural links, so whatever credit they passed is lost.
The disavow tool is for a narrow case: Google says most sites won’t need it, and it exists for unnatural links that have caused, or are likely to cause, a manual action. John Mueller has said the idea of toxic links is “made up by SEO tools”. Link attributes are hints, too: since 2019 Google has treated nofollow, sponsored and ugc as hints, so a nofollow link can still help a page get found. Paid placements must still be qualified.
What AI search changes
Google says there are no extra requirements for AI Overviews or AI Mode beyond being indexed and eligible for a snippet. So what helps a page rank still helps it get retrieved; what changes is which off-site signals line up with being mentioned.
Ahrefs’ study of 75,000 brands found that branded web mentions correlated with AI Overview visibility at 0.664, branded anchors at 0.527 and backlinks at only 0.218. A December 2025 follow-up found YouTube mentions correlated most strongly (0.737 for ChatGPT) and described backlinks and referring domains as very weak. Seer Interactive found page-one Google rankings tracked brand mentions in GPT-4o answers, while backlinks had a weak or neutral effect.
These are correlations; they don’t show that building mentions causes citations. Rankings also overlap less with AI Overview citations than they did. In March 2026 Ahrefs found 37.9% of cited pages ranked in the top 10, down from 76.1% in July 2025, and puts it down to query fan-out.
So a backlink strategy is now also a brand mention strategy. In my experience the most valuable links come with a mention in a relevant context: an industry article naming you as the source, a tool roundup, a conference recap. They help with Google, and they’re the kind of mention that AI systems can pick up as citations. If brand authority is your focus, the course AI Search & LLMs: Entity SEO and Knowledge Graph Strategies for Brands covers how a brand becomes a recognisable entity across the web, and how LLMs co-cite explains why it matters who you’re mentioned alongside.
What makes a quality backlink
My old guide rated “follow” status as critical. Here’s how I’d rate the signals now:
| Signal | Weight | How to check |
|---|---|---|
| Relevance of the linking page and site | High | Read the page. Embeddings can score topic match at scale |
| Editorial placement in the main content | High | Article body, or footer, sidebar, widget? |
| Independence (not your site, a network or a paid placement) | High | Who runs the site? Was the link given or bought? |
| A real audience | Medium | Readers and an engaged community. Traffic figures are third-party estimates |
| Natural, descriptive anchor | Medium | Brand, URL or a plain description chosen by the author |
| Link attribute (follow, nofollow, sponsored, ugc) | Low to medium | A hint. A nofollow link from a well-read, relevant page still brings visitors and a mention |
| Third-party scores (DA, DR, Trust Flow, Authority Score, DataForSEO rank) | Triage only | Good for sorting a list. Google doesn’t use them |
Moz says Domain Authority isn’t a direct Google ranking factor and is meant for comparison, so I’ve dropped the “be careful below DA 40” rule I used to give; it never had a source. DataForSEO’s rank uses the original PageRank formula with a damping factor of 0.5: a reasonable popularity proxy, and still a third-party estimate.
The AI-powered backlink workflow
Six steps, with mlforseo.com’s data from 7 November 2026 as the example.
1. Audit what you have
DataForSEO reported 2,611 backlinks to mlforseo.com from 611 referring domains (288 main domains), a domain rank of 219 out of 1,000, and a backlink spam score of 34 against 3 for the site itself: the spam is pointed at the site from outside. 743 backlinks came from .xyz domains alone, and 1,117 (about 43%) from five cheap TLDs (.xyz, .website, .space, .online, .site).
Through MCP this took a few plain-English prompts and a few cents of API calls. The no-code tutorial has the prompts, and the Python notebook does the same with logged costs. New to the API? Start with DataForSEO API in Python.
2. Judge quality with context
Sorted by rank, my top referring domains were my personal site (983 links, mostly sitewide navigation), my newsletter archive host, then real sites mixed with course-listing and course-resale sites. The editorial links I value most, from iPullRank, Oncrawl and Moz, sat further down with modest scores. A rank-sorted list buries exactly the links you want more of.
This is where AI earned its place. Once I told Claude which domains I own, it separated owned links from earned ones, grouped more than 1,100 anchors as link-seller spam from their wording alone, and flagged that the high-rank links weren’t the valuable ones. The audit guide covers the signals, and an LLM-as-a-judge setup labels every link with a reason you can check.
3. Ignore spam you didn’t build, clean up what you did
More than 1,100 of my backlinks were nofollow ads for a Telegram link seller, and 26 followed links advertised a PBN service with my domain in the anchor. Nobody at MLforSEO bought these, and Google’s systems are built to ignore exactly this kind of link, so I’m doing nothing. Disavow is for links you or an agency built in breach of the policies, ideally after trying to get them removed. What to do after the audit walks through that decision.
4. Reclaim what you’ve lost
Google drops links that point at a 404 (the link goes nowhere, as Mueller put it), so broken backlinks are value lost to your own URL changes. DataForSEO flagged 33 broken backlinks. Claude noticed all 33 pointed at two URLs (30 from one article translated into 30 languages) and checked the live status codes first. Both returned 200: they had returned a 404 at some point when DataForSEO’s crawler visited, and the index hadn’t caught up. Nothing needed fixing.
Lesson: link data lags, so check the live status before building a redirect or emailing anyone. When a URL really has moved, a 301 to the closest live page fixes it, and the 404 and redirect mapping notebook speeds up the matching.
5. Prospect with a relevance filter
A link gap lists sites linking to your peers but not to you. Against two peers I got 10 domains, mostly junk (an app download page, a thesis-writing service, a domain valuation site) and three genuine prospects: an SEO blog, an education platform and a developer community. Use three to five real peers, and filter for relevance with embeddings or an LLM reading page titles before anyone writes an email.
6. Earn links worth earning
AI can’t do this step. Links come from having something worth linking to: original data, a free tool, a clear explanation nobody else has written, an expert view a journalist needs.
| Step | Where AI helps | Where it doesn’t |
|---|---|---|
| Audit | Pulls and summarises data through MCP or an API in minutes | The data is a third-party sample and can be weeks old |
| Judge quality | Groups anchors, spots patterns, explains each label | Doesn’t know what you own unless you tell it |
| Clean up | Separates links you built from ambient spam | Can’t see whether Google neutralised a link; Search Console only reports manual actions |
| Reclaim | Groups broken links by target, checks status, suggests redirects | Will trust a stale “broken” flag unless asked to check |
| Prospect | Scores relevance, researches sites and people | Can’t make an irrelevant site relevant |
| Earn | Helps analyse data and draft, edit and repackage content | Can’t create expertise, data or relationships |
Types of backlinks and how to earn them
Editorial links
A writer links to you because your page helped them make a point, like the iPullRank, Oncrawl and Moz links in my profile. These count most. Earn them with content that offers something new (a method, a dataset, a worked example), and make sure people writing about your topic see it.
Resource and tool links
Free tools, templates, notebooks and checklists are easy to cite, so tutorials, roundups and courses link to them. The 30 translated-article links above all point at one guide in our resources library. Keep resource URLs stable: my broken-link scare came entirely from moving resources.
Original data and studies
Publish your own data with the method, and make charts easy to reuse with a credit line. Skip infographic directories (low-quality directories count as link spam); original visuals still earn links when embedded with attribution.
Expert quotes and digital PR
HARO, the go-to in my old guide, closed in December 2024 under its later name, Connectively, and the brand was sold on in 2025. I’d take the slower route: be quotable on a narrow topic, answer journalists quickly, and join roundups, panels, podcasts and events where your view adds something; their pages and recaps often link back.
Guest posts, with honest caveats
I used to call guest posting an amazing backlink source. Now I’d treat it as an audience tactic. Google names optimised anchors in guest posts as link spam, and the July 2021 link spam update warned that unqualified guest-post links can lead to action. Write for publications whose readers you want, and keep any link natural, such as in your author bio.
Mentions, reclamation and communities
Ask for a link where a site already mentions your brand; a plain mention passes no PageRank, though it may still count for AI visibility. Ask for credit where your original images are used, and offer your page when a niche resource someone links to has gone.
Forum posts, comments and social profiles should carry ugc or nofollow, which is fine: the visitors are the point. Partners, associations and testimonials for tools you actually use can also produce relevant links, with paid relationships marked sponsored.
Anchor text: types and best practice
Anchor text is the clickable text of a link. Five types show up in any profile:
- Exact match: the keyword itself, such as “backlink strategy”.
- Partial match: the keyword plus other words.
- Branded: your brand name, such as “MLforSEO”.
- Naked URL: the address, such as “mlforseo.com”.
- Generic: “click here” or “read more”.
Google’s link best practices ask for descriptive, concise, natural anchors without keyword stuffing, and the spam policies treat optimised anchors in placed content as link spam. Google has never published a safe exact-match percentage, so ignore anyone who quotes one. Where you have a say (a bio, a partner page), ask for your brand name or a plain description. Branded anchors had the second-highest correlation with AI Overview visibility in the Ahrefs study.
Anchors also expose spam. The top anchors to mlforseo.com fell into three groups: my own navigation, honest brand and URL anchors, and link-seller ads that gave themselves away by their wording.
Monitoring your backlink profile
I check monthly on active sites and quarterly on quiet ones:
- New and lost referring domains, more than raw backlink counts (one sitewide footer link adds hundreds of backlinks). Use at least one third-party index; each sees a different slice of the web.
- Where new links sit, what the linking pages are about, and sudden spikes of commercial anchors.
- Broken backlinks, with a live status check before acting.
- Mentions with and without links, including YouTube and AI answers. My AI Overview tracker helps with the last part, and consistent brand details make mentions easier to attribute to you.
- The Search Console Links report: free and straight from Google, but a sample, with tables limited to 1,000 rows, no nofollow data, and sometimes removed links.
A backlink outreach checklist for 2026
My old checklist, updated. The rule for AI: use it to research and personalise, never to send. An email generated for 500 sites reads like one.
- Find relevant sites, newsletters, communities and creators with engaged audiences: relevance and readership first, scores second.
- Run a link gap against three to five real peers, filter it for relevance and read every prospect.
- Note where peers write as guests or get quoted, and pitch ideas only you can write.
- Ask for links on unlinked mentions and credit links on your reused images.
- Offer replacements for broken links in your niche when your page is a genuine fit.
- Contribute to roundups, expert panels, podcasts, events and webinars, and ask organisers to link your talk page.
- Publish free resources and tools, and tell the people who would use them.
- Keep your most-linked pages current, so they stay the version worth linking to.
- Build relationships with partners, associations and local organisations, and respond to mentions on social platforms, review sites and forums.
- Let AI summarise a prospect’s recent work and draft a first version. Then rewrite it yourself, check every claim, and send one at a time.
Start with the audit if you’ve never looked at your profile, or the Claude and MCP tutorial if you want your own numbers this afternoon.
FAQ
Do backlinks still matter for SEO in 2026?
Yes. Google uses links as a relevancy signal and to discover pages, though its staff say they matter less than they used to. A few relevant editorial links beat volume.
Do backlinks help you appear in AI Overviews and ChatGPT?
Indirectly: a page has to be indexed and competitive to be retrieved, and links help with that. In Ahrefs’ 75,000-brand study, brand mentions correlated far more strongly with AI visibility than backlinks (0.664 against 0.218 for AI Overviews). It’s a correlation, so treat it as a direction.
Should I disavow toxic backlinks?
Usually no. Google says most sites won’t need the disavow tool, and SpamBrain neutralises spam links. Disavow links that you or an agency built in breach of the spam policies, especially if you have a manual action.
Is AI-written outreach a good idea?
For research and first drafts, yes. For sending at scale, no: mass AI outreach is spam with better grammar, and it costs you goodwill with the editors you want.
MLforSEO Academy
Take your skills further — pick your path
Semantic ML-enabled Keyword Research
by Lazarina Stoy
Query understanding, search intent and semantic clustering for how people and AI really search.
View course →AI Search Optimisation & Agentic SEO
by Beatrice Gamba
How agentic systems retrieve, evaluate and select — and how to become the source they cite.
View course →AI Search, LLMs, Entity SEO & Knowledge Graph Strategies
by Beatrice Gamba
Build the entity authority and knowledge-graph presence that makes AI recognise and trust your brand.
View course →Introduction to Machine Learning for SEO
by Lazarina Stoy
Ship practical ML workflows — classification, clustering, entity extraction — no CS degree required.
View course →


