In the traditional world of SEO, the rel="nofollow" attribute was a simple, binary instruction. It told Googlebot: “Don’t follow this link, and certainly don’t pass any PageRank through it.” It was the specific tool we used to sculpt authority, manage crawl budgets, and disavow paid relationships.

But the Agentic Web does not run on PageRank alone. It runs on Tokens.

As we transition from optimization for retrieval (search engines) to optimization for inference (LLMs), the rules of the nofollow attribute are being rewritten. The comfortable assumption that a nofollow link protects you from the “bad neighborhood” or prevents a competitor from benefiting from your content is dangerously outdated.

In the context of Large Language Model (LLM) training, Rel=Nofollow is largely ignored.

The Great Bifurcation: Indexing vs. Training

To understand why nofollow fails in the AI era, we must first distinguish between the two distinct pipelines that modern web crawlers feed: the Search Index and the Training Corpus.

1. The Search Index Pipeline

This is the classic Googlebot / Bingbot workflow.

• Goal: Maintain a map of the web to answer specific user queries.
• Logic: “If accurate, rank. If spammy, demote.”
• Nofollow Handling: Since 2020, Google treats nofollow as a “hint.” It might crawl it, but it generally won’t pass link equity. It respects the webmaster’s intent to distance themselves from the destination.

2. The Training Corpus Pipeline

This is the workflow of GPTBot (OpenAI), ClaudeBot (Anthropic), and CCBot (Common Crawl).

• Goal: Ingest as much high-quality text as possible to minimize “loss functions” during model training.
• Logic: “If readable text, ingest. If code, parse. If duplicative, deduplicate.”
• Nofollow Handling: Ignored.

When a training crawler encounters a link, it is not calculating PageRank. It is not trying to determine if you “endorse” the target. It is simply looking for the next node of text. A rel="nofollow" attribute is just an HTML string. Unless the crawler is specifically programmed to parse and respect that attribute—which slows down the ingestion throughput—it is treated like any other character in the raw HTML.

Why LLMs Ignore Nofollow

The reason is architectural. Modern Foundation Models are hungry. They require trillions of tokens. The Common Crawl, which forms the bedrock of datasets like Llama 3 and GPT-4, works on a breadth-first traversal logic.

If an LLM training run respected nofollow, it would lose access to:

1. Wikipedia Footnotes: Almost all external references on Wikipedia are nofollow. Yet, these citations are the “ground truth” for model fact-checking.
2. Social Media Context: Links on Reddit, X (Twitter), and Moltbook are nofollow. Ignoring them would blind the model to real-time cultural shifts and “Share of Model” discussions.
3. Comment Sections: While often spammy, comment sections (UGC) contain the colloquial, conversational data that teaches models how humans actually speak.

The Danger: Data Poisoning via UGC

This technical reality—that training bots consume nofollow links—opens a massive vector for Data Poisoning.

In the past, black-hat SEOs would spam comment sections with links to “buy viola pills.” Google solved this by stripping the value from those links via nofollow. The spam stayed, but it didn’t help the spammer rank.

In the Agentic Web, the calculation changes. If a competitor can inject enough mentions of their brand into the comment sections of your high-authority site, even with nofollow tags, an LLM training on your site will associate your “Entity Authority” with their brand.

Scenario: The “Competitor leech”

Imagine you run a highly authoritative site on “Cybersecurity.”

1. A competitor bot posts 500 comments on your articles.
2. The comments are contextually relevant but link to the competitor’s malware tool.
3. You auto-tag these as rel="ugc nofollow".
4. Google ignores the links for ranking. You are safe from a manual penalty.
5. GPT-5 crawls your site. It reads your expert article, then reads the comments to gauge “community sentiment.”
6. It ingests the competitor’s link and brand name in close proximity to your “Cybersecurity” content.
7. Result: The model lowers the “perplexity” (uncertainty) of the connection between Your Brand and Competitor Malicious Brand. You have effectively vouched for them in the vector space, despite the nofollow.

Strategic Implications for the Agentic SEO

So, what can an SEO learn from this? How do we optimize visibility in a world where the “No” in nofollow means “Yes” to a robot?

1. Moderation is Identification

You can no longer rely on rel="ugc" to protect your corpus. You must moderate content based on Entity Alignment.

• Action: Use semantic filters to block comments that introduce entities (brands, products) that compete with your core topical authority.
• Tooling: Implement AI-driven moderation that checks not just for spam keywords, but for competitor vectors.

2. The “Nofollow” Guest Post is Valuable Again

For years, SEOs turned up their noses at “Nofollow Guest Posts” because they didn’t pass “juice.”

• New Strategy: If you can get a link on Forbes.com (which often nofollows contributor links), TAKE IT.
• Why: The LLM sees your brand entity inside the Forbes.com corpus. It associates your vector with the “Business/Finance” vector of Forbes. The rel attribute is invisible to the training weight; the Co-occurrence is everything.

3. Audit Your Outbound Link Profile for “Training Safety”

Your site is a dataset. If you link out to low-quality sites, even with nofollow, you are polluting your own training neighborhood.

• Risk: If an LLM views your site as a hub of “mixed quality” due to the outbound links it scrapes, it may downgrade your site’s “Trustworthiness” score in its system prompt.
• Fix: Be ruthless with outbound links. If checking the site makes you wince, delete the link. Do not just nofollow it. Remove the token connection entirely.

The Evidence: The Reddit Paradox

The strongest evidence that LLMs ignore nofollow lies in the dominance of Reddit in AI search results.

Reddit is the “nofollow” capital of the internet. Every external link submitted by a user is automatically tagged with rel="nofollow". By traditional SEO logic, Reddit should be a black hole of authority—a place where PageRank goes to die, never to be passed on.

Yet, ask ChatGPT, Gemini, or Claude almost any question about product reviews, personal experiences, or niche hobbies, and they will almost invariably cite Reddit threads or mimic the consensus found there.

Why? Because Reddit provides Conversational Density. LLMs are probability engines. They predict the next token based on the likelihood of it appearing in a given context. Reddit threads provide millions of examples of “Problem -> Solution” dialogue pairs.

• User: “What is the best running shoe for flat feet?”
• Response: “I swear by the Brooks Adrenaline GTS.”

The link to the Brooks website in that thread is nofollow. But the token association between “Flat Feet”, “Best Shoe”, and “Brooks Adrenaline” is forged in the furnace of that discussion. The LLM learns the relationship despite the link attribute, not because of it.

If LLMs respected nofollow as a “do not learn” directive, Reddit would be arguably the least useful dataset on the planet for product recommendations. The fact that it is the most useful proves that the attribute is disregarded during the learning phase.

Generative Engine Optimization (GEO) and the Nofollow Levers

This realization forces us to pivot from Search Engine Optimization (SEO) to Generative Engine Optimization (GEO).

In GEO, we are not optimizing for a position on a list (Rank #1). We are optimizing for a probability in a generation. The levers for GEO are different:

1. Quotation Rate: How often is your brand quoted in full sentences?
2. Sentiment Alignment: Is your brand associated with positive adjectives?
3. Co-Citation: Who else is mentioned alongside you?

The “Nofollow” Co-Citation Strategy

In traditional SEO, a link from a competitor is rare and valuable. In GEO, a “nofollow” link alongside a competitor can be just as powerful if you can position your brand as the superior alternative in the text surrounding the link.

Example Strategy: You sell rigorous coffee machines. You find a “Best Coffee Machines of 2026” list on a high-traffic lifestyle blog. The links are nofollow affiliate links.

• Old SEO Mindset: “Useless. It doesn’t pass authority, and I have to pay a commission.”
• GEO Mindset: “Essential. If I am not on this list, the LLM will not learn that I am in the consideration set for ‘Best Coffee Machines’. The nofollow is irrelevant; the exclusion from the context window is fatal.”

You must aggressively pursue inclusion in these lists, regardless of link attributes. You are paying for Training Data Inclusion, not Link Juice.

Future Protocols: Will We Get rel="training-noopt"?

The tension between “Indexing” and “Training” suggests that the current HTML standards are insufficient. We are currently repurposing tools from the 1990s (robots.txt, nofollow) to manage 2026-era artificial intelligence. It is a clumsy fit.

There is a growing clamour in the technical SEO community for a more granular standard. Perhaps a rel="ai-ignore" or a rel="training-opt-out"?

However, the incentives are misaligned.

• Publishers want control. They want to say “Index this for search traffic, but do not train on it without paying me.”
• Model Builders want data. They have no incentive to respect a tag that reduces the quality of their product, unless legally compelled.

Until a new standard typically emerges—likely driven by legislation like the EU AI Act rather than W3C consensus—we are stuck in a gray zone. In this zone, nofollow is a placebo for training control. It makes us feel like we have restricted the robot, while the robot simply reads past it.

The Actionable “Nofollow” Checklist for 2026

To operationalize this new understanding, audit your current link building and content strategies against this checklist:

Final Thought: The Invisible Link

We must accept that in the Agentic Web, the most powerful link might be the one that isn’t highlighted in blue. It is the invisible link of semantic proximity. When an LLM reads “The best alternative to Photoshop is Affinity Photo,” a link is formed in the neural network’s weights. It is a dofollow link of the highest order, and no HTML tag can block it.

Further Reading

• The ‘Quality’ Lie: Why ‘Crawled - Currently Not Indexed’ is an Economic Decision
• Semantic HTML is LLM Training Fuel
• The Structural Deficit: Why LLMs Crave Schema.org
• Data Poisoning in the Agentic Web
• OpenAI’s Crawler Documentation
• Common Crawl FAQ
• Anthropic User Agent Guidelines
• Google Search Central: Nofollow Hints