SparkBeacon asks ChatGPT, Claude, Perplexity and Google's AI Overviews your customers' real questions, repeatedly, and reports how often each one names you, with the uncertainty, per engine, and what to do about it.
Named in AI answers, per engine and never averaged. 983 answers, 4 runs pooled over 28 days; the 95% intervals are in the table below.
| Questions | 16 on the panel · 6 invisible · 3 cited · 7 named · 0 recommended |
| Per engine | ChatGPT 3.3% (1.7% to 6.3%) · Claude 6.1% (3.7% to 9.8%) · Perplexity 8.8% (6.1% to 13%) · Google AI Overview 5.6% (3.1% to 9.7%) |
| Sources | 96% earned · video #7 on Google's answers |
| Movement | 0 confirmed · 3 too early to call |
Being cited, being mentioned, and being recommended are three different events with three different owners. Most tools measure the second and sell it as the fourth. SparkBeacon measures each rung separately, because the work to climb each one is different.
A crawler reaches and parses the page. If the firewall turns it away, nothing below this line can happen.
An engine uses the page as a source and links it. The page was good enough to quote.
The brand name appears inside the sentence the reader sees. The cheapest rung to fix and the most often skipped.
The model shortlists the brand for a stranger's unbranded question. Earned sources put a name there.
Recognition is not recommendation. In one test of 112 startups, 99.4% were recognised when named; 3.32% surfaced in unbranded discovery.Reverse-engineered
Single-author thesis (arXiv 2601.00912), not peer-reviewed. Direction matches our own ledgers; the exact figures are one sample.
Cited is not mentioned. One of our own brands: cited six times, named zero, competitor recommended in the same answer. The fix was a sentence.Our data
SparkBeacon ledger
The mention is where the reader stops. Click-through 8% with an AI summary, 15% without; 1% clicked a link inside the summary.Proven
Pew · 68,879 searches · 900 US adults · March 2025
A single visibility score hides four different problems. SparkBeacon is structured so that every question sits on a rung and every rung has its own evidence and its own next step.
The site audit presents each AI crawler's own user-agent and records whether the firewall serves it. A reporter on the brand's server then logs the visits that identify themselves as those crawlers, by name and not yet by address, so the log is a floor. Where a firewall verifies bots by network address, the audit says unverifiable instead of guessing, because a request wearing a crawler's name from anywhere else is supposed to be refused.
| GPTBot | served |
| OAI-SearchBot | served |
| ChatGPT-User | served |
| ClaudeBot | served |
| Claude-User | served |
| Claude-SearchBot | served |
| PerplexityBot | served |
| Googlebot | served |
| Bingbot | served |
| Applebot | served |
| Meta AI | served |
| Bytespider | served |
| CCBot | served |
Every source an engine used is kept and marked as the brand's own or someone else's. A channel or profile on a platform the brand does not own counts as its own page too, so a video library is credited from the first video. The earned share is measured for this brand, not quoted from an industry average.
Each question is asked several times per engine and every answer is read for the brand's name. The result is a rate with a 95% interval, because five answers can differ from the next five by sixty points. The gap between cited and mentioned is copy work: the name goes inside the sentence, never beside it.
| ChatGPT | 3.3% · 1.7% to 6.3% · 8 of 245 |
| Claude | 6.1% · 3.7% to 9.8% · 15 of 246 |
| Perplexity | 8.8% · 6.1% to 13% · 26 of 294 |
| Google AI Overview | 5.6% · 3.1% to 9.7% · 11 of 198 |
| Cited, not named | 111 of 487 times the brand's pages were used |
The panel is built from unbranded questions in the customer's own words: who should I list my house with, which agent is best at marketing a luxury home. Who gets named instead is listed, question by question. A rung crossing only counts as a move when the two windows' intervals stop overlapping, in either direction.
AI answers change from one ask to the next. A tool that asks once and reports a percentage is reporting noise with a decimal point. These are the rules every SparkBeacon figure passes before it is shown.
The 95% interval on a single answer. It cannot tell 10% from 80%.
Still wide. Recent runs of the same question pool together, and at about 35 answers the interval reaches ±16, without paying for it twice.
A blended score describes an engine that does not exist. Every figure names its engine.
Intervals: Wilson score, 95%. Study figures from the sources cited in The Visibility Standard v2.1; nearly every large study in this field is vendor-run and the Standard applies its own test to each.
After the first answer, a searcher asks one of five things next. Each is a different page. SparkBeacon maps the customer journey from real search demand and Google's own what-people-ask-next chains, then sorts every follow-up into its type so the missing page is visible before anyone has asked for it out loud.
A type with no live follow-up is a page nobody has written for this journey yet. It is the cheapest page to be first on, because the demand is already visible in the chains and no incumbent is answering it.
Live from Anne Sostman's journey map · sorted by wording, free, deterministic
Part of understanding a field that changes monthly is knowing which of last quarter's tactics have already been tested and failed. SparkBeacon will not charge you for any of these.
97% of 137,000 domains' files got zero requests (Ahrefs, May 2026); Google does not read it. We generate one because it is free, and say so on the file.
Controlled test: AI Overview citations −4.6%, AI Mode +2.4%, ChatGPT +2.2%. Required as a commerce mechanism, not a citation lever.
One number across engines describes no engine. Every SparkBeacon figure names its engine.
69% of the citations they earn appear in answers recommending a competitor.
Below seven answers the interval is wider than the claim. We pool, or we say too early.
Reddit is 20.8% of the top-50 external citation domains in one B2B software sample and 1.1% of citations in a realtor sample. Different denominators, and a nineteenfold gap. Measured per brand, never inherited.
The API path was retired in August 2025.
OpenAI's caught 26% and false-flagged 9%; 61% false-positive on non-native English. Filler is penalised, not assistance.
The Visibility Standard treats the mechanics as constant and the brand as eight variables. Setting them in order is the whole onboarding; a blank dial is usually the finding.
One name, address, phone, licence and domain, byte-identical everywhere.
In the customer's words, never the trade's.
Where the buyer actually is.
One person, chosen A or B on every dimension. Both is not an answer.
Measured, not assumed: which sources the engines draw on for this brand.
The evidence this market believes: photos, numbers, video, reviews.
The one true before, from the brand's own past.
Which regulator, which rules, and the named person who decides.
In order: fill the dials · verify rung one · publish the ungated evidence layer · commission the first proof · set the entity string everywhere · baseline under M · only then choose channels · name the compliance owner
Every measurement is stored forever, so any before-and-after is a query rather than a memory. These are the modules, each with its own append-only history.
Your questions, four engines, repeated, with intervals. A stopped run resumes without paying twice.
Every question on its rung, pooled over a window, with confirmed and too-early moves.
Baseline against now: what is working, what is not, competitor movement with how and why.
Stages from search demand, the follow-up chains, the five follow-up types.
Who the engines name on the questions where you never appear.
Cited domains, earned versus owned, where video ranks on your own questions.
Firewall answers per crawler, and a server-side reporter for the visits that identify as each crawler.
Crawl, AI readiness, issues with copy-paste fixes, a free llms.txt.
Measured positions from live results pages, plus the licensed index's view.
Where you are listed, where details differ, which searches show a map box.
Auto-discovered rivals, why their authority differs, how to overtake it.
What runs on your searches; where page one is weak enough to take.
Humans arriving from AI, with the referrer, plus branded-search demand.
What your package includes and where you are this month, on one page.
Each package sets how many questions are on the panel, how many answers are sampled per question per engine, and how many runs a month. The limits are on the settings page, month to date, so there is never a surprise. Pricing is not published yet; the launch list hears first.
How many of your customers' questions are asked, across four engines, every run.
How many times each question is asked per engine. More answers, tighter interval. Recent runs pool, so depth compounds.
Weekly or monthly, plus the ones you press yourself. Every run is kept forever.
Every package includes the full instrument: ladder, journey, sources, crawler access, site audit, rank tracking, listings, competitors, progress report. The package changes how much is measured, never what you can see.
One email when SparkBeacon opens. Nothing else.