¶01 What happened, and what it is worth
On 7 September 2026 a Google AI Overview answered "marketing agency for NetSuite partners" by naming three firms. Artifexa was the first of them, and the citation panel beside the answer linked two Artifexa pages.
That is a single observation of a single result, on one day, in one session, and search results are personalised and localised. It is not a ranking, it is not a study, and it is not repeatable evidence of anything on its own. Recorded honestly, its label is observed: seen directly, once, in a stated state.
It is still worth writing about, for one reason. Almost everything published about being cited by AI is written by firms that have not been, reasoning from the outside about a mechanism nobody outside Google or OpenAI can actually see. This page reasons from the other direction: here is a firm that was named, here is what its cited pages have in common, and here is which of that is likely to matter.
¶02 Why an assistant will not recommend most firms
An answer engine is not choosing a favourite. It is assembling a sentence it can stand behind, and it will not put a claim in that sentence if it cannot source it.
Most professional services sites give it nothing to work with. They are written in adjectives. Trusted, leading, end-to-end, results-driven. None of that can be repeated by a machine that is answering a stranger's question, because none of it says anything a reader could check. Strip the adjectives out of a typical consultancy homepage and you are usually left with the company name and a list of service categories.
The second problem is corroboration. A firm whose entire evidence base sits on its own domain is asking to be taken on trust, and the same cross-referencing that catches an inflated claim also catches a contradicted one. If your site says one thing and your directory listing, your LinkedIn page and your client's site say something slightly different, the safest thing an engine can do is not mention you.
The third is identity. If a machine cannot tell reliably which company you are, it will not name you in case you are somebody else. There is a real firm with a name one letter from Artifexa's operating in an adjacent sector in another country. That ambiguity is a competitor for your own name, and it is solved with structured data rather than copy.
¶03 The unit is the sentence, not the page
This is the part that changes how you write, and it is the difference between search and answer.
Search asks which page best answers a query, so a page can win on the strength of the whole. An answer engine writes the answer itself and lifts sentences it can attribute, so a page can rank perfectly well and still be useless to quote, because every sentence on it depends on the one before.
Test your own copy on it. Take any sentence that makes a claim, put it on its own, and ask whether it still means something to someone who has read nothing else. "We deliver measurable outcomes for our clients" fails, and it fails everywhere. "The review is free, written, and yours to keep whoever does the work" survives, because it says what a thing is, what it costs and what happens next.
The same test applies to numbers. A figure with no period attached is not a measurement, it is a decoration, and an engine that quotes it inherits your sloppiness. "1,200 visitors" is unquotable. "925 unique visitors between 1 September 2025 and 31 August 2026, from site analytics" can be repeated by anyone without risk, which is exactly why it gets repeated.
¶04 What the cited pages had in common
Two Artifexa pages were cited. Neither was written for this. Both were written to answer a question a buyer actually asks, which may be the whole finding.
What they share, and what most competitor pages on the same query do not:
A question as the title, answered in the first paragraph
Not a service name. The page that ranked first is called "Who does marketing for NetSuite partners in the UK?" and its opening line says there is no approved list. The answer is above the fold and does not require the rest of the page.
Named alternatives, including the ones that are not us
That page describes four kinds of provider and is honest about what each is good and bad at. A page that only argues for itself is a sales page, and a sales page is a poor source. A page that fairly describes the options is a useful one, and it is far more likely to be the thing an engine leans on when a reader is comparing.
Figures with periods and sources attached
Every measured number on the site names where it came from and when it was taken, and the ones that could not be measured are labelled as estimates or as not measured at all.
An organisation a machine can identify
Structured data naming the firm, its address, its contact point and the places it appears elsewhere. An llms.txt file stating which pages are cornerstone. Neither is difficult and both are commonly absent, including on sites selling this exact service.
Prices on the page
Published rather than gated behind a call. It is a fact rather than a claim, and facts are what a source is made of.
¶05 What this does not prove
Longer than the section above it, deliberately.
It does not prove causation. These pages have those properties and this firm was named, and there is no way from outside to show the second followed from the first. Something else may be doing the work, including the simple fact that few firms write about this niche at all.
It does not prove durability. AI Overviews are regenerated, they differ between users and locations, and a result observed once may not be there next week. Anyone quoting a single citation as a permanent position is doing the thing this page is warning about.
It does not prove commercial value. Being named is not an enquiry. Artifexa has not yet measured a single enquiry that arrived because of an AI answer, and will say so until it has.
And it is not a method you can buy with certainty. Nobody outside the companies running these systems knows how sources are chosen, and the honest version of this advice is that the properties above are good practice for a human reader too. If the answer engines changed tomorrow, a page that states its claims plainly, sources its numbers and describes its competitors fairly would still be the better page.
¶06 What to change on your own site
In order, cheapest first.
Date every number and name its source. This costs an afternoon and it is the single change that most improves whether a claim can be repeated. An undated figure is unusable to a careful reader and to a machine alike.
Rewrite your five most important claims so each survives on its own. Cover the page and say the claim out loud. If it needs the paragraph above it to make sense, it is not going to be quoted.
Answer the question your buyer actually types, using their words as the heading, with the answer in the first paragraph rather than at the end. Most service pages bury the answer under an argument for the service.
Mark up who you are. An organisation with an address and a contact point, so nothing has to guess which firm you are, and so a similarly named company somewhere else does not take the mention.
Then be findable somewhere that is not your own domain, so a claim on your site can be checked against something that is not your site.
None of that is a trick and none of it expires when the models change. If you want the same ground measured on your own firm rather than described in general, the free Visibility Review does exactly that, with every finding labelled the way the observation at the top of this page is labelled.