How AI assistants decide which companies to recommend

The three-step decision

Strip away the interface and every assistant answers a "recommend me a vendor" question the same way.

Step one: retrieve. The assistant turns the question into searches, often several at once, and pulls back a shortlist of pages: listicles, directories, vendor sites, forum threads. If nothing about you is retrieved, you are out before anything intelligent has happened. Retrieval runs on search indexes, which is why classic SEO plumbing still gates everything.

Step two: corroborate. The model weighs what it retrieved. Do multiple independent sources describe this company consistently? Does the vendor's claim about itself match what others say? Is the information current? Companies that appear once, describe themselves one way while the internet says another, or last got written about in 2019, get discounted here.

Step three: compose. The assistant writes its answer from the surviving material, quoting or paraphrasing the passages that were easiest to lift. If your site states plainly what you do, for whom, and where, your own words often become the description. If it does not, the machine composes your description from whatever else it found, and you live with the result.

Where each assistant looks

The mechanics above are shared; the retrieval sources are not. Citation overlap between engines is small, roughly one answer in ten, so it pays to know who reads what.

Two practical consequences. First, submit your sitemap to Bing Webmaster Tools; it takes ten minutes, it feeds ChatGPT and Copilot, and in India essentially nobody does it. Second, use Perplexity as your test bench: publish, wait a week, ask again, and you have real evidence of whether your content is citable.

The corroboration step is where Indian SMEs lose

Most founders I speak to assume the fix is on their website. Usually the website is the healthiest part of their footprint. What is missing is the agreement layer around it: the category listicles that rank for "top X in India" and never mention them, the directory profiles with a defunct address, the review platforms with three reviews from 2020, the total silence on Reddit and Quora where their buyers ask questions.

Assistants read that layer as the jury. A company whose website says one thing while the surrounding record is thin, stale or contradictory does not get recommended; it gets hedged, or skipped. The unglamorous work of getting listed, described consistently and reviewed recently moves recommendations more than another website rebuild ever will.

What a recommendation-worthy footprint looks like

• Your site answers "what do you do, for whom, where" in its first screen of text, and the same one-line description appears on LinkedIn, in directories, and in your structured data.
• You appear in at least a few of the category pages that rank for your buyers' "best X" searches, because those are what assistants harvest for shortlists.
• Recent, real reviews exist somewhere an assistant can read.
• Something about you is dated within the last year: a news mention, an updated page, a published article. Currency is corroboration too.
• Nothing machine-readable contradicts anything else. Consistency is cheap and its absence is expensive.

What to do this quarter

Baseline first: run the query panel exercise and write down what each assistant says about you and your category today. Then fix in this order: your own site's one-line clarity, your directory and LinkedIn consistency, one or two category listicle inclusions, review recency. Recheck monthly. Expect Perplexity to move first, Google surfaces in weeks, ChatGPT's description of you last; the ordering is about how each engine retrieves, and it is covered in the pieces linked below.