Anchor text vs. mention context: what really teaches the AI who you are

You're optimizing the clickable words of the links pointing to your site — but for the AI that's the wrong detail. What really matters is the text of the fifty to one hundred words around your name every time someone mentions you: those are the words the AI uses to build its description of who you are. If those words don't speak to your positioning, the AI describes you incorrectly in every answer. Rewriting the context of those mentions — often with small tweaks to already-published copy — is enough to change how the AI presents you to potential clients.

Take one of your recent backlinks. Look at the anchor text. Now look at the full sentence surrounding the anchor. It’s that sentence — not the anchor — that the AI uses to understand who you are.

For years SEO taught us to optimize two words inside a link. To show up in AI answers that game still counts, but much less than before. What matters is the paragraph around your brand name: the 50-100 words that precede and follow the citation are the raw material the model uses to decide what you are, for whom, and when to cite you.

In my articles on Digital PR for AI I keep coming back to this point because it’s the most underrated one among Italian professional firms. Today I’ll explain it with an A/B comparison I ran on two notaries in Romagna.

What an AI model sees when it finds a mention of your brand

An LLM doesn’t read the web like a Google crawler. It doesn’t “weigh the anchor” as a ranking signal. It reads sequential text and, as it reads, it compresses information onto a few anchor tokens that it will later use to answer.

In the world of LLM research this mechanism is documented. Pang et al. describe it like this:

“In a recent study, Wang et al. (2023a) demonstrate that label words in prefix demonstrations can act as anchors during inference, providing an effective context compression approach for improving inference efficiency in in-context learning.”

Pang et al., 2024

Translated: during inference, the model concentrates the context’s information onto a few anchor tokens. These aren’t HTML links, they’re semantic keywords that capture the meaning of what surrounds them.

The operational consequence is direct for your firm or your company. If your brand appears in a paragraph packed with “corporate transactions,” “mergers,” “Ravenna,” those words become the semantic anchors your name stays hooked to. If it appears in a generic paragraph (“a local notary firm”), your name doesn’t tie to anything specific.

Anchor text alone is no longer enough

Here’s something important to understand. In old-school SEO, the anchor “corporate-transactions notary Ravenna” pointing to your site was half the job done. In the AI-answers ecosystem that link, on its own, carries much less weight.

Again from the Pang et al. paper:

“Additionally, Pang et al. (2024b) observe that LLMs tend to attend to only a few, yet consistent, prefix tokens during inference. However, the specific tokens utilized are often unpredictable and uncontrollable.”

Pang et al., 2024

It follows that the model chooses on its own which tokens to treat as anchors, and the ones it picks are “often unpredictable and uncontrollable” from the outside. You can’t force ChatGPT or Perplexity to treat your optimized anchor as the center of gravity of meaning. What you can do is saturate the surrounding context with the words you want attached to your name.

It’s the same principle I explained when talking about backlinks as a citation proxy: the link stays useful as a signal of the page’s reliability, but the informational value for the model lies in the text around it.

The A/B test I ran: two notaries in Romagna

A few weeks ago I took two notary firms specialized in corporate transactions, both based between Ravenna and Forlì. I’ll call them Notary A and Notary B (names redacted on request).

Comparable profile:

• Same sector: notarial practice with a focus on M&A, capital contributions, corporate restructurings
• Similar number of backlinks: 47 (A) and 43 (B) from quality Italian sources (Il Sole 24 Ore, Italia Oggi, local legal outlets, chambers of commerce)
• Comparable domain age: 12 and 14 years

The difference lay in the context of the mentions, not in the links.

Notary A: their backlinks came from articles where the firm’s name appeared in dense paragraphs — “the corporate merger handled by firm A in the Ravenna industrial district,” “the conversion of a joint-stock company into a limited-liability company assisted by notary A for an agricultural cooperative in Romagna.” The surrounding paragraph always carried 3-4 technical specialization terms plus the location.

Notary B: same types of publications, but the mentions were “as firm B explained to us” or “according to notary B of Ravenna.” No operational context around them.

I ran 15 queries on ChatGPT and Perplexity, split like this:

• 5 generalist queries: “best notary for corporate transactions in Romagna”
• 5 specific queries: “notary for a merger between two cooperatives in the province of Ravenna”
• 5 sector queries: “who to turn to for a business contribution in Emilia-Romagna”

Result (indicative test, not a controlled study): Notary A cited in 11 answers out of 15. Notary B cited in 2 out of 15. Same links, different contexts, AI visibility 5x.

The sample is small and AI answers vary between sessions. But the pattern matches what I’ve observed over the past few months across 20+ professional firms in north-eastern Italy: the context of the mention predicts visibility in AI answers better than the number of backlinks.

How to check the context of your mentions in 15 minutes

You don’t need an expensive tool to get started. Three steps are enough.

Open Google Search Console and filter the brand queries. Export the external pages you receive from the “Links” section. For each external page, do something trivial: read the paragraph where your name appears. Not the anchor, the whole paragraph.

Then open displaCy ENT, paste that paragraph, and look at which entities the parser recognizes. If the paragraph around your brand contains entities consistent with your positioning (places, professional categories, types of service), the context is working for you. If it contains disconnected entities or no meaningful entity at all, your name is floating in a semantic void.

Binary thresholds, so you can decide right away:

• The brand paragraph contains at least 2 keywords of your positioning → useful mention
• The paragraph contains only your name with no operational context → neutral mention, almost invisible to the AI
• The paragraph is about something else and mentions you in passing → inert mention

For a serious analysis across hundreds of mentions you need professional tools, but this entry-level check already tells you where you stand.

The mistakes I notice most often

Asking the journalist only for the right anchor. “Can you link me with ‘corporate-transactions notary Ravenna’?” is the wrong request. What you need is: “can you write a sentence explaining what I worked on?” The anchor is then chosen by the editor.

Press releases with the brand name repeated in empty paragraphs. “Firm X announces. Firm X is pleased. Firm X confirms.” The model learns that you’re an entity that announces things, not that you’re a specialist in business contributions.

Company profiles full of adjectives, devoid of specializations. “A dynamic, innovative outfit, rooted in the local area” creates no semantic anchors. “A notary firm with 15 years of experience in extraordinary transactions and trusts” does.

Never checking the context of the mentions you receive. I have clients who pay four-figure PR fees without ever having read a single paragraph of how they were mentioned.

What to do starting tomorrow

1. Audit the last 20 mentions of your brand and classify each paragraph as useful / neutral / inert
2. In your next PR brief to journalists, ask for a sentence of operational context, not just an anchor
3. Review your press office boilerplate: it must contain 3-4 explicitly named specializations
4. Compare the context of your mentions with the 3-5 competitors the AI already cites in your sector — copy the semantic structure, not the words

The link on its own stays useful, don’t throw it away. But to show up in AI answers the real work is on the paragraph. The 50-100 words around your brand are the crash course you give the model on who you are: if you saturate them with the right entities, the AI learns; if you leave them empty, you haven’t spent your PR money badly — you’ve just invested in a signal the AI can’t read.