How to improve your AI search visibility

Almost every company we audit is surprised by its own number, in one direction or the other. Guessing wastes months, so start by measuring.

Write ten questions a buyer would actually type. Not keywords, questions. Cover five shapes, because the shape changes the outcome more than anything you publish: the broad category question, alternatives to a named competitor, two named competitors compared directly, a use case question, and pricing.

Put each question to more than one engine. Claude, ChatGPT, Perplexity and Google AI Overviews disagree with each other more than they agree. A single-engine number is a reading, not a position.

Score two separate things per answer. Whether your company was named, and whether your domain was cited as a source. These diverge more than people expect. In our data 29.4% of companies were cited more often than they were named, and nine were named in exactly zero answers while their own site kept appearing in the citation list.

One trap to avoid, because we fell into it ourselves. If four of your ten questions contain a competitor's name, that competitor will appear in those answers almost every time. Counting that as a mention measures your own question, not the engine. Score name-seeded questions separately from open ones. When we split our own category data this way, two of the five agencies we were tracking dropped from thirteen mentions to zero.

Before publishing anything new, check what the engines already think is true about your company. This is the cheapest work available and it is usually the most wrong.

Three things go stale and keep circulating. Ownership, when a company has been acquired, carved out or renamed and third-party profiles still carry the old parent. Category, when a directory files you under something adjacent to what you actually sell. And name, when your product has several variants in circulation and the engines treat them as different entities.

None of that is a content problem. It is a data problem on pages the engines already read, which means it is correctable by contacting the source, and it moves in weeks rather than months.

On your own domain, three things do the same job. One canonical company name used consistently. Organization structured data with a sameAs array listing every property you own, so an engine can tell your LinkedIn page, your GitHub organisation and your research all belong to one company. And the same identifier referenced across every page, rather than each page declaring an unlinked company of the same name.

This is the single most common gap we find, and it has a distinctive signature.

Question shape changes naming rates sharply. Best-of questions named a company 41% of the time in our data. Direct comparison questions named one 20% of the time. And the pattern shifts with how visible a company already is: companies named in none of ten answers mostly lose broad category questions, while companies named in seven to ten lose head-to-head comparisons 68.8% of the time instead.

The reason is usually visible on the company's own site. A page will name three competitors and then say "see how we compare, request a demo". That hands the engines the rival names and withholds the answer, so the comparison gets written from rival and third-party pages, without you in it.

Write the comparison properly instead. Name both sides. Concede rows honestly rather than sweeping every column, because a page that admits a limitation is the most citable thing in most categories and almost nobody publishes one. Include the cases where the other product is the better choice.

An answer engine asked to name the best software in a category needs names to return. A page written to rank for a keyword often contains none.

We audited a company whose main guide page was fifteen hundred words on how to choose software in its category and did not name a single company, including its own competitors. It said at one point that everyone knows the biggest players, and then declined to say who they were. That page was cited zero times in a hundred and eighty citations, while six competitor pages that answered the same query with an actual list of platforms were cited repeatedly.

The opposite failure exists too, and it is subtler. Another company published a category guide that ranked itself first, with the best score in every row of its own comparison table. The engines cited that page three times and named every competitor in the table except the company that wrote it. A vendor ranking itself reads as a claim. The competitor names on the same page are the part that is not self-serving, so that is the part that gets used as evidence.

The lesson from both: name companies, and do not make yourself the answer to every row.

Across the hundred most-cited domains in our study, 77.4% of citations went to vendor-published material. Review platforms took 10.2%, media 6.5%, analyst firms 5.9%, and community sites including Reddit 0.0%. That last number surprises people, so it is worth being precise: this measures what the engines cited when answering B2B software questions in this dataset, not whether Reddit matters generally.

The practical reading is that vendor-written content dominates the citation layer, and most of it was written by your competitors. The authority is also thin. 56% of cited domains appeared exactly once, and the ten most-cited domains together accounted for only 12% of all citations. There is no settled reference layer in most categories, which is the whole reason the position is still available.

So the work is to be one of the vendors whose material gets read. Answer the questions your category leaves unowned. In our data, use case questions returned no vendor at all 40% of the time, which means nobody has claimed them. Publish original data if you have any, because it is the most citable asset a company can own and almost nobody publishes it. And correct your entry on the third-party pages that already rank, because those are being read whether you like them or not.

Re-run the same questions on the same engines. Not similar questions, the same ones, or you are measuring the change in your question set.

Expect entity corrections to move first, in weeks, because they are data fixes on pages that already get read. Expect answer-layer work to be slower, because the pages have to be found, read and trusted before they change anything.

Report the range, not a point. If you tell a board that you went from three mentions to eight, the first question should be how much that number moves on its own. If you have not measured that, you do not know.

And separate the two scores. Being cited more is progress even when naming has not moved yet, because it means your material is being read. Being named without being cited means the engines are describing you from somewhere else, which is a different problem with a different fix.