How AI Assistants Decide Which Cambridge Firm to Cite

A buyer types something like who supplies cryogenic control electronics near Cambridge. The assistant does not read the whole web live. It works from material it has already processed. In many cases it adds a retrieval step over a live search index before it answers.

The answer is then synthesised from whichever sources the system judges most reliable for that exact question. Reliability here is not a score anybody can see. It is mostly the product of agreement between independent sources.

Three things have to happen in sequence. Your pages have to be retrievable. A claim about your company has to be extractable from those pages. That claim then has to survive a comparison against everything else the system holds about you.

1. Crawler access, checked directly

   Your robots.txt has to allow GPTBot, PerplexityBot, ClaudeBot, and Google-Extended. Several site platforms block those agents by default, and a firm can sit invisible for a year without anyone noticing. This is the cheapest failure on the list to fix.

2. A claim in plain language

   A model needs one sentence that states what the company does in words a non-specialist uses. A page that opens with a description of the underlying physics gives the system nothing to lift. The plain sentence and the technical detail can both live on the page, with the plain one first.

3. Agreement across independent sources

   Your site, your Companies House record, your LinkedIn page, and any press coverage all describe the same company. A model that finds three different descriptions of your category has a reason to hedge or to say nothing. Consistency is a ranking factor in the AI world in a way it never quite was in the Google world.

4. An unambiguous entity

   The company name has to resolve to one organisation. A model needs enough surrounding context to separate your firm from a similarly named business elsewhere, and from the university whose technology you may have licensed.

Cambridge carries a name collision that most UK cities never face. Cambridge in Massachusetts is a dense technology and biotech cluster in its own right. A model asked about a Cambridge company therefore has two plausible cities to resolve against before it says anything at all.

The University adds a second layer. A spinout that licenses university technology is often described in press coverage under the university's name rather than its own trading name. The research is attributed correctly and the company is barely mentioned. Years later the model has plenty of material about the science and almost none about the business.

The naming habit adds a third layer. Plenty of firms here carry Cambridge inside the company name itself. A system then has to separate the place from the brand, and the shorter the rest of the name is, the harder that separation gets.

1.Silicon Fen is the common nickname for the technology cluster around Cambridge, and it is useful shorthand rather than a formal boundary.

You control four things directly. You control crawler access, the wording of your own claims, the structured data that carries those claims, and the consistency of your description across the profiles you own. Everything else is influence rather than control.

That distinction matters for how the work is scoped. A retainer that promises citations is selling something nobody can deliver. A retainer that fixes crawler access, rewrites the capability statement, corrects the schema, and aligns the public profiles is selling work that is real and checkable.

The AI visibility page covers the discipline in general terms. This page sits inside the Cambridge deep-tech field guide, which holds the rest of the cluster.

Ask the engines the questions your buyers ask, and write down what comes back with the date attached. A citation baseline is a dated table rather than an impression. Without the date the exercise is worthless, since these systems change underneath you constantly.

Use real buyer questions rather than your own company name. An investor rarely starts with a name they have never heard. They start with a category question, and the name is what they hope to receive back.

Repeat the same questions after the fixes ship. The comparison between two dated runs is the only honest evidence available in this field, and it is the reason every check on this site carries a re-check date.

It does not claim that these four signals are the whole mechanism. No AI platform publishes its selection criteria, and the weighting between factors is not public. The signals described here are the ones that are observable, controllable, and consistent with what the platforms themselves document about crawling and sourcing.

It also does not claim that a fix produces a citation. A dated baseline, a specific list of changes, and a fair re-check are what can honestly be offered.