How AI assistants decide which brands to recommend.

Two terms describe the same shift from two ends. Both matter, and neither is enough on its own.

They depend on each other. An engine cannot recommend what it cannot read, and a perfectly readable page that no engine chooses to cite is not visibility. Work on one without the other and the number you report will not move the way you expect.

Search-grounded assistants — ChatGPT search, Perplexity, Google AI Overviews and others — answer a buyer’s question in roughly four moves:

1. 1The question. Often a shortlist or comparison: “best tools for…”, “X vs Y”, “which vendors are SOC 2 compliant”.
2. 2Retrieval. The engine fetches pages it is allowed to crawl: your site, your competitors’, and third-party pages such as reviews, forums and publications.
3. 3Composition. It writes an answer that names some brands, orders them, and sometimes steers the buyer away from one.
4. 4Citation. It links to some of the sources it leaned on — which are frequently not the brands’ own sites.

Most visibility tools count how often your brand is named. But an answer can name you and advise against you in the same sentence. Counted as a mention, that answer raises your score while it loses you the buyer.

The fix is to grade how each answer treats you, and let that grade decide what the mention is worth. This is the scale Vidrys uses:

Alongside that grade, four numbers describe where you stand:

Share of voice

The share of recorded answers that name you at all. Necessary, and on its own not enough.

Endorsement

How the answer treats you once you are named, on the scale above.

Position

Where you land in the answer: first pick, alternative, or footnote.

Citation share

Which domains the engine leaned on — yours, a competitor’s, or a third party’s.

Because answers vary, a visibility number is only as good as the method behind it. Five rules keep it honest:

• Ask the questions buyers ask. Category, comparison and requirement prompts — not your brand name, which every engine already knows.
• Draw each prompt more than once. A second draw separates a real move from run-to-run variation.
• Put an interval on the number. Report a change only when it clears the confidence interval, and call smaller moves inconclusive.
• Version the scoring. When the formula or judge changes, stamp the version and never trend scores across versions — or last quarter’s report quietly becomes wrong.
• Wait for enough evidence. Don’t call a head-to-head against a competitor until you have seen you both in enough answers. Vidrys waits for eight co-appearances.

Answer Engine Optimization comes down to three questions, in the order sites usually fail them.

When a competitor wins an answer you wanted, start with why. Read the answer the way a buyer does and look for the edge that decided it. The usual ones:

Then make the move that closes that specific edge:

• Publish the page that answers the question directly — often a comparison or use-case page.
• Earn a place on the third-party sources the engine cited for that question.
• Refresh pages that are losing on recency, with dates and explicit figures.
• Remove technical blockers first; they cap every other fix.

Finally, measure the same prompt again, over enough draws to clear the interval. A fix you cannot see move the answer is a guess, not a result.

Comparing platforms that do this? See how Vidrys compares.