How we measure AI visibility
In one paragraph
ModernTechLap asks your tracked questions on the AI systems you connect, stores every answer with the engine, the model that produced it and the date, and reads mentions, recommendations, position and cited sources off that answer text deterministically. The AI Visibility Score is 50% mention rate, 30% recommendation rate and 20% position. Runs are repeated on a schedule because answers vary, and every number can be traced to a dated answer you can read.
Most vendors in this category will tell you a number and not how they got it. A visibility score you cannot recompute is an assertion, so here is the whole method, formulas included, along with the parts of it that are weak.
How a measurement is produced
You confirm the questions before anything runs
Tracking starts from the questions your buyers actually ask, not from keywords. The product drafts a set from what you sell, who you sell it to and which competitors you name, and you edit, add to or delete from it before the first run. Only the questions you keep active count against your plan limit.
Each run asks every active question on the systems you connect
A run submits your active questions to each AI system you have connected, and stores every answer in full alongside the engine, the exact model id that produced it, the citations it gave and the timestamp. Failed provider calls are retried, and an answer that never arrives is recorded as missing rather than counted as an absence.
Mentions are read off the answer text deterministically
Whether you were mentioned, recommended, in which position, and which sources were cited is derived from the stored answer by string analysis, not by a second model. The same answer always produces the same verdict, so a trend line can only move when the answers move. Every field is something you can check by reading the answer, which is printed next to it.
Runs repeat on your plan's cadence, and each one is dated
Because answers vary between runs, one run is a sample and the series is the signal. Runs repeat monthly, weekly or daily depending on your plan, you can trigger extra ones up to a daily cap, and every point on every chart is stamped with the date and the run it came from.
Completed actions are re-measured about a week later
When you mark an action done, the questions that action targets are asked again roughly a week afterwards, and the before and after numbers are shown side by side. It is an honest comparison of two dated samples, not an attribution claim: the page says so where the numbers sit.
How many questions, and how often
There is no universal right number of questions, so the plan sets a ceiling and you decide how to spend it. Repetition matters more than volume: the same questions asked again on a schedule is what turns a snapshot into a measurement.
| Plan | Questions | Refresh | Runs a day |
|---|---|---|---|
| Free | 25 | Monthly | 2 |
| Growth | 300 | Weekly | 3 |
| Scale | 1,500 | Daily | 4 |
| Agency | 7,500 | Daily | 6 |
Questions are per brand. The daily cap covers scheduled and manual runs together, because each run is real model calls on every connected system. Compare the plan limits.
How each number is calculated
Every metric is arithmetic over stored answers. Nothing is estimated by a model, and nothing is smoothed, weighted by a secret factor or adjusted after the fact.
AI Visibility Score
0 to 100100 × (0.5 × mention rate + 0.3 × recommendation rate + 0.2 × position score)
Half the score is how often you are mentioned, three tenths is how often you are recommended, and a fifth is where you sit when you do appear. The position score is the average of 1 divided by your position across the answers that name you, so first place counts 1, second 0.5 and third about 0.33.
Mention rate
percentanswers naming you ÷ all answers in the window
A name counts when the answer uses your brand name, your domain or the registrable part of it. A shortened or wrong form of your name is recorded as it was written rather than normalised away, because how a model holds your entity is itself a finding.
Recommendation rate
percentanswers recommending you ÷ all answers in the window
Recommended means named among the first three vendors in the answer without a negative qualifier beside it. The qualifier test is a keyword check, which is why the answer is shown next to the verdict and you can disagree with it.
Share of voice
percenttimes you were named ÷ times any tracked name was named
Your share of every mention of a tracked name, yours and your competitors', across the window. It moves when a competitor gains ground even if your own mention rate is flat, which is the point of having it.
Position
rankorder of your first mention among the tracked names in one answer
Counted within a single answer, over the tracked names that answer mentions. An answer that names you without a readable ordering contributes nothing to the position score rather than being guessed at.
Cited sources
count per domainanswers citing a domain, split by whether the domain is yours
Recorded when the AI system returns its citations. When it returns none, the answer is marked as uncited instead of having sources inferred for it.
Engine, model, date and language controls
These four things change the answer more than anything else, so they are recorded on every answer and held steady between runs. A result reported without them is not reproducible even in principle.
- Engine
- All 7 systems are available on every plan, and none of them is a paid upgrade: ChatGPT, Claude, Gemini, Google AI Mode, Perplexity, Grok, Microsoft Copilot. You choose which ones a question runs on, and every metric can be read per engine, because they disagree with each other more than they agree.
- Model
- Each system is called through its own API with web search enabled where the provider offers it. The exact model id is stored with every answer and shown beside it, so a jump in a chart can be checked against a model change rather than explained away as one.
- Date
- Every answer carries the timestamp of the run that produced it. Charts are built from dated runs, comparisons name both dates, and a partial run is labelled as partial instead of being averaged in silently.
- Language and market
- Questions are asked exactly as you wrote them, in that language, from a neutral context with no account history and no location personalisation. One market at a time: running the same question set as different countries is not supported yet, so a result is about the neutral context and not about a specific country's buyers.
Known limitations
Every measurement has an edge. These are ours, and they are here rather than in a footnote because a buyer who finds them later is entitled to wonder what else was filed under small print.
Answers are not deterministic, so a single run is one sample
Ask the same question twice and the shortlist can differ. That is a property of the models, not a fault in the measurement, and it is why runs are repeated and dated, why charts are series rather than single figures, and why we will not put one run on a slide as a ranking.
Absence in your tracked questions is not universal absence
If you do not appear across your 300 tracked questions, that is evidence about those 300 questions. It is not proof that no AI system ever names you. The product words findings that way and you should read them that way.
Detection is a readable heuristic, not a judge
Mentions and positions are matched from text, and the recommended-or-not test leans on a list of negative cues. It separates a top pick from a passing mention well enough to trend, and it will occasionally be wrong on an unusual sentence. The answer sits next to the verdict for exactly that reason.
Personalisation is outside our control
These systems vary their answers by account history, region and staged rollouts. Holding the run conditions steady makes our comparisons meaningful between runs. It does not mean your buyer, signed in on their own account, sees what we saw.
We do not have the models' selection logic
Nobody outside those companies does. We can show what changed, when, and which sources the answers leaned on. We cannot give you the weighting that produced it, and any vendor who offers you one is guessing.
Re-measurement is a comparison, not an attribution
The before and after around a completed action is two dated samples of the same questions. Model updates, competitor moves and ordinary variance all sit inside that gap. Where a change cannot be attributed, the product says it cannot be attributed.
What we do not guarantee
We do not guarantee that any AI system will mention, recommend, cite or include your brand, on any plan, and we will not sell you a package that implies otherwise. The models belong to other companies, their selection logic is not published, and their output changes between runs. ModernTechLap measures and influences the evidence around a business; it does not control AI model outputs. What you get is a measurement you can recompute and a prioritised list of changes that are worth making on their own merits.
Common questions
How is the AI Visibility Score calculated?
How many questions does ModernTechLap track?
Why does the same question give different answers?
Are mentions detected by another AI model?
Does ModernTechLap guarantee that AI will recommend my brand?
See what the method produces
The sample baseline shows every section of this page applied to an illustrative company. Or run 25 questions on your own brand for nothing.