AI Search (GEO visibility)

AI Search tracks your visibility inside AI answer engines - Perplexity, Google's AI answers (Gemini) and ChatGPT - the new front page that increasingly sits in front of traditional search.

How it works

1. Prompt basket - maintain a set of questions/keywords to test (the platform suggests a starting set from what the Brain knows; add or remove your own). Capped at 30 active prompts - see Basket size and scan limits below.
2. Scan - the platform asks those prompts across the engines, automatically each week and on demand.
3. Results - a headline metrics band shows how visible you are at a glance, with every number carrying its fraction (e.g. cited in 14/27 answers) so you can always see what it's counted from.

Reading your results

Every scan sends each prompt to each engine, and each of those is one answer (e.g. 9 prompts x 3 engines = 27 answers checked). Within an answer, your brand can be:

• Top pick - you're the answer's main recommendation.
• Cited - the engine links to your site as a source. This is the strongest practical signal: it means the engine both trusts your content and sends readers to it.
• Mentioned - your brand is named somewhere in the answer, but without a link.
• Absent - the answer doesn't include you at all.

Engine cards

One card per engine, leading with that engine's citation rate - the share of prompts where it cites your site - plus the underlying fractions (n/n prompts cite you · n/n mention). The progress bar is colour-coded: teal at 60%+ citation rate, amber at 30-59%, red below 30%. If one engine's card is red while the others are teal, that's where your visibility gap is.

Visibility score

A single 0-100 score summarising your AI visibility, shown as a ring with a band: Excellent (80+), Good (60-79), Fair (40-59) or Needs work (below 40). It's a fixed formula - 40% AI presence + 40% citation rate + 20% top-pick rate - calculated directly from your scan counts (no AI judgement involved), so it moves only when the underlying results move. Use it to track direction month over month rather than as a target in itself.

If a scan didn't get through your whole basket, the score is withheld rather than shown with a caveat - see Incomplete scans below.

KPI tiles

Four tiles under the score, each with its fraction:

• Prompts tracked - how many questions are in your basket, and the total answers checked.
• AI presence - the share of answers that mention your brand at all. The broadest measure: "do the engines know we exist?"
• Citation rate - the share of answers that cite your site as a source. The one to grow: citations drive clicks.
• Top pick - the share of answers where you're the recommended option. The hardest to earn and the most valuable.

Share of Voice

A ranked list of every brand the engines named across your prompts - you and your competitors. Each row shows the brand's share % of all brand mentions, its mentions/total fraction, and how often it was cited. Bars are scaled to the leader, so the gaps between ranks are easy to read. Being outranked here tells you which competitor the engines reach for first when answering your customers' questions.

Prompt-by-prompt matrix

A grid of every prompt against every engine, with a colour-coded dot per answer (the legend sits in the card header): solid teal = Top pick, teal ring = Cited, amber = Mentioned, grey = Absent. A "Cited by n/3" column counts how many engines cite you for that prompt, and rows are sorted best-first - so your strongest prompts are at the top and the ones needing content work are at the bottom.

Once any prompt has been cited at least once, a First cited column appears showing the date of the earliest scan in which an engine cited you for it. That is your evidence of what changed and when.

Visibility over time

Every scan is kept, so the tab charts your visibility score, mention rate and citation rate across your scan history, oldest first. This is the answer to "are we becoming more visible", as opposed to "are we visible today".

A few things worth knowing about how it is measured:

• Each point is measured against that scan's own answer count, so a scan where an engine was unreachable cannot read as a visibility drop.
• A scan that produced no usable answer at all, or that didn't get through the whole basket, is left off the chart rather than plotted as a drop that never happened. The card tells you when that has happened - see Incomplete scans below.
• The line appears from your second scan. Before that you get the single reading and its date. Engines re-crawl over weeks, so scans about a week apart read best; a day-to-day wobble means very little.
• Per-answer detail is kept for 400 days, which covers a quarter-long programme plus a year-on-year comparison. The scan-level scores are kept beyond that.

Incomplete scans

A scan asks every active prompt of every connected engine. If it can't get through the whole basket in one run - the basket grew a lot, or an engine was slow - the run ends marked incomplete rather than quietly reporting what it managed as a finished scan.

When that happens:

• An amber banner says so, and tells you how many answers were covered out of how many were expected.
• The visibility score is withheld. A score calculated on part of your basket isn't comparable with one calculated on all of it, and a headline number gets quoted and compared whatever the small print says.
• The reading is left out of the trend, and the chart tells you how many scans were left out. Plotting a partial run next to full ones would invent a dip that never happened.
• The answers that were collected are still saved and still shown in the prompt matrix. Nothing is thrown away; it just isn't scored.

Just run the scan again. If it keeps stopping short, the basket has outgrown a single run - pause the prompts you care least about, or split the work across two scans.

Which lever moves a prompt

Not every prompt you are missing from is a content problem, and the tab does not treat them as if they were. Each prompt is classified by what an answer engine actually draws on when it answers that kind of question, and the recommended action follows the classification:

• Problem / how-to and definitional questions ("why doesn't my ad platform match my analytics", "what is marketing intelligence") are won by whoever explains it best. These appear under Answer-content opportunities, with a one-click hand-off to the Copywriter.
• Comparison questions ("X vs Y", "alternatives to X") are won by both your own comparison page and third-party listicles, so they get the answer-content action too - but publishing is only half of it.
• Vendor-selection questions ("best agency for X", "top tools for Y") are assembled from directories, review sites and roundups, so a page on your own site will not win one. These are listed separately, under Uncited - but answer content is the wrong lever, with the reason stated, and the work belongs in the citation-gap section below.
• Brand questions about your own company are answered from your entity footprint - llms.txt, structured data and third-party profiles - rather than from a new article.

The classification is shown as a label against each prompt in the basket; hover it for the reasoning.

Citation gaps

A citation gaps card shows where competitors get cited but you don't - with the sources worth pursuing to close the gap. This is where vendor-selection prompts get won.

Helping the engines find you

The platform supports the technical groundwork too (such as an llms.txt and structured data), making it easier for AI engines to understand and cite your site.

Basket size and scan limits

Every prompt is a real query to every connected engine, so a 30-prompt basket on three engines is 90 live searches each time it runs. Three limits keep that predictable, and all three are shown in the tab before you hit them.

30 active prompts. Paused prompts don't count towards it, so pausing is the tidy way to make room - the prompt stays in your list with its history and can be brought back whenever you want. The basket header shows where you are (18 / 30 active) and the add field tells you when you're full. The cap also keeps a scan finishing in a single run, which is what stops scans coming back incomplete.

If Suggest from brain produces more prompts than you have room for, it adds what fits and tells you exactly how many it left out. It won't quietly fill you to the ceiling.

8 on-demand scans a month. That's the Scan now button. The counter sits next to it, and it resets on the 1st.

Your weekly scheduled scan is not affected by any of this. It runs regardless of how many on-demand scans you've used. That matters: the weekly scan is what builds your trend over time, so running out of manual scans never costs you the measurement.

One hour between on-demand scans. Answer engines re-crawl over days and weeks, so scanning again ten minutes later returns the same answers at full cost. The button tells you when the next one is available.

Availability & usage

AI Search is available on every plan. Scans draw on a usage allowance, counted in engine queries - one per active prompt per connected engine - see Usage & limits. Findings feed the Brain, so visibility trends shape your reports and recommendations.