How every number here is measured.
claw.mobile publishes three kinds of measurements: the monthly AI recommendation index, the daily price tracker, and per-business AI visibility scans. This page is the canonical description of how each one works, in enough detail that you could rebuild the numbers yourself. If a behaviour is not described here, we do not do it.
Methodology, November 2026. Everything below describes the systems as they run today. When the mechanics change, this page changes with them.
- Index cadence
- Monthly, fixed set
- Price checks
- Daily, verified
- Scan gate
- 85% or held
Short answer: At a glance
claw.mobile runs three measurement systems. The AI recommendation index asks the same 40 versioned buyer questions to ChatGPT, Claude, Perplexity and Grok every month with one neutral prompt and classifies each answer against a fixed roster; failed calls shrink the denominator and partial runs are labeled. The price tracker checks official pricing pages daily and never lets a failed extraction overwrite verified data. The $99 AI Visibility Scan generates 25 niche questions per buyer, runs them on the same four engines, and only finalizes when at least 85% of expected answers were classified. Rankings are not for sale and every published number traces to stored raw answers. As of November 2026, this page is the canonical reference for all of it.
The AI recommendation index
Which tools the AI assistants recommend when real buyers ask. Published monthly on the index page with the raw dataset attached.
A fixed, versioned question set
The index asks 40 buyer questions, written the way a real buyer types them: category queries ("best AI app builder"), head-to-heads ("Bolt vs Lovable"), jobs to be done ("build an internal tool from a spreadsheet") and per-tool checks ("is Lovable worth it"). The set is frozen under a version label, currently v1. Changing any question text bumps the version, so months under the same version are always comparable. Buyer-intent categories were layered on in August 2026 as metadata on the same frozen set: no question changed, the version did not bump.Four engines, one call each, neutral prompting
On the first of each month, every question is asked once per engine through the official APIs: ChatGPT (OpenAI), Claude (Anthropic), Perplexity, and Grok (xAI). Default parameters, no temperature overrides. The only system prompt is "Answer as you normally would.", identical for every engine, so none is steered toward or away from any vendor. The exact model that answered is recorded per run and named on the index page.Extraction against a fixed roster
Each answer is classified by a separate model call against a fixed whitelist of tools. "Mentioned" means the tool is named at all. "Recommended" means the answer presents it as a pick for the asker: it tops a list, is called best for the use case, or is explicitly suggested. A tool named only in passing, or only to warn against it, counts as mentioned and never as recommended. Names outside the roster are ignored, so a hallucinated product cannot enter the index.Every number audits back to a raw answer
One database row is stored per engine and question with the raw answer text, so every percentage on the public page can be traced to the exact answers behind it. The monthly aggregate is published as a static JSON file under a CC BY 4.0 license, linked from the index page.Failures shrink the denominator, never the truth
Every engine call gets up to three attempts. A question that still fails is recorded as a failure and excluded from that engine's denominator, so percentages are computed only over answers that actually arrived. A run with any failures is marked partial in the published data and on the page.
The price tracker
Current pricing for the tracked AI app builders and coding tools, checked daily against each tool's official pricing page and published on the changelog.
Daily checks against the official source
Once a day the tracker fetches every tracked tool's official pricing page. A plain fetch is tried first; pages that render their prices with JavaScript fall back to a headless browser. A page only counts as fetched when the rendered text has real length and at least one dollar figure, so an error page or cookie wall never masquerades as pricing.Extraction is validated before it counts
Extracted plans must pass a gate before they are accepted: between 1 and 12 plans, every plan named, at least one numeric price, and no price outside 0 to 3,000 dollars a month. Anything that fails is treated as a failed check, not as data.Verify before stamp: failed checks never overwrite good data
When a fetch or extraction fails, the previous verified snapshot is kept and the tool is flagged as degraded. On a first run with no good snapshot yet, the entry is seeded from the pricing data the site already publishes and labeled as seeded until a live scrape succeeds. The published data always says which of the three states each tool is in.Change events only between verified states
A price change ("Replit Core $25 to $20"), added plan, or removed plan is only reported when both the previous and the current snapshot were verified against the live page. A failed scrape can never produce a phantom change. The raw current state is public at /data/tracker/latest.json and the feed renders on the changelog.
The claw.mobile score
The number the home page ranking is ordered by. It is built only from figures published on this site, so any row can be recomputed.
One board for each kind of tool
AI app builders, AI coding assistants and no-code builders are ranked separately. Nobody chooses between Lovable and GitHub Copilot, so one list of all of them answered no real question.Two inputs, fixed weights
Our own rating carries 60% of the score and AI recommendations 40%. The rating is taken as a share of 5. AI recommendations are the times the tool was recommended in the month's index, as a share of the most-recommended tool on the same board. The rating leads because it is the only input that says whether the tool works; AI recommendations say how well known it is.No test, no score
A tool we have not tested has no score and no rank. It is listed under the tested tools, in order of AI recommendations, and marked "Not yet tested". We do not score it on one input and we do not give it an average in place of a rating.Price sits beside the score, not inside it
The cheapest paid plan is shown in its own column. Folded into a score, one unusually cheap plan rescales every other tool of its kind, and a reader can no longer tell a good tool from a cheap one.What the AI figure cannot tell you
Each assistant is asked each question once, and assistants rarely give the same list twice, so a gap of a few answers between two tools is within the noise. The figure measures how often a tool is recommended, not how good it is.Awards
Most recommended by AI: The tool AI assistants recommended most often in its category in the month’s index. Not given when two tools tie. Top rated in our tests: The highest rating we have given in the category. Given only where we have rated at least two tools, so it is never a contest of one. Best value: Among tools we rated 4 out of 5 or higher, the one with the lowest-priced paid plan; the higher rating wins a tie. Given only where we have rated at least two tools.Nothing in it can be bought
No input is touched by a sponsor spot, a directory listing, an award badge, an affiliate programme or priority testing. Where we earn a commission on a tool, the board says so and the data download marks it. Every input is in ranking.csv on the open data page, so any position can be recomputed.Changes to the method
Version 2.0, 18 November 2026: one board per kind of tool; price moved out of the score; a tool without a rating no longer gets a score from its other inputs. The same day, the ratings for Framer, FlutterFlow, Glide and Adalo joined the no-code board, when their review pages began stating them. Version 1.0, 18 November 2026: a single list scored on AI recommendations (40%), our rating (35%) and price (25%), replaced the same day because it scored tested and untested tools on different formulas.
The AI Visibility Scan
The same measurement, pointed at one business and its niche, is what our sister product Saymetry runs. The rules below are the ones both share.
Questions are generated per scan, from the buyer's own market
Each scan builds a set of 60 questions: around 24 core buyer questions, including up to 3 the buyer wrote themselves (used verbatim, never rephrased), plus around 36 long-tail probes across specialty, style, budget tier and location whose job is finding pockets where the buyer already shows up. The buyer's site is fetched read-only (homepage plus up to three obvious subpages) to ground the set; nothing is inferred beyond what the site's own text supports.Same engines, same machinery as the index
Every question runs across the same four engines through the same call machinery the monthly index uses, imported rather than copied, so the two measurements can never quietly drift apart. One deliberate difference: scan answers run with web search enabled, because that is how the consumer apps answer buyer questions; when an engine's search variant fails, the plain call stands in and the answer honestly records that no search happened. Answers are classified under a per-scan roster: the buyer's business, their competitors, plus open capture of any other name the answers recommend. The buyer appears in their own leaderboard even at zero.The 85% completeness gate
A scan may only finalize when at least 85% of all expected answers were classified. Below that threshold the scan is held: nothing is stored as final, the buyer is emailed nothing, and the run is retried. This gate exists because a partial scan once looked deliverable, and a report built on a fraction of its answers is worse than a late one.One free re-run, and failure cannot destroy a report
Every scan includes one free re-run: the buyer replies with adjusted questions (up to 6, used verbatim) and the scan runs again, publishing under the same report link, labeled as a re-run of the original date. If anything fails mid-way, the original report is left untouched. Orders are claimed atomically, so two overlapping runner passes can never both fulfil the same scan.
Recommended, mentioned, cited
Three words that get blurred together everywhere else. Here they are three different measurements, and mixing them up changes the numbers, so they are pinned down once.
Mentioned
The tool or business is named anywhere in the answer, in any role. Being mentioned costs nothing and proves little; it is counted because the gap between mentioned and recommended is itself informative.Recommended
The answer presents the name as a pick for the asker: it tops a list, is called best or great for the use case, or is explicitly suggested. A name that appears only in passing, or only with a warning attached, is mentioned, not recommended. Share of voice on this site always means share of answers that recommend, never share of answers that mention.Cited
A domain that appears in the sources an engine attaches to its answer. In scans, every engine runs with web search enabled, so cited domains are counted across all four; in the monthly index, which deliberately stays on plain calls so its methodology only changes between months, sources come from Perplexity, the one engine that browses by default. Cited measures where answers come from; recommended measures what they say.
What we never do
Rankings are not for sale
No payment changes a position in the index, the price tracker, or a scan report. Partners buy intelligence about the numbers and presence on editorial surfaces; the numbers themselves come from the monitors, and the top tools get their index badge free whether or not they ever pay us.Failed checks never overwrite good data
The rule is the same everywhere: the tracker keeps the last verified snapshot when a scrape fails, a scan below the completeness gate is held rather than delivered, and a failed engine call shrinks the denominator instead of being guessed at. Silence is recorded as silence, never filled in.No fabricated numbers
Every published figure traces to stored raw material: a raw answer row for the index, a verified page snapshot for the tracker, a classified answer for a scan. If the raw material does not exist, the number is not published, and partial runs are labeled partial.No steering
Engines are asked with default parameters and one neutral system prompt, identical across engines. We never prompt an engine toward or away from any vendor, including the ones that pay us.Everything paid is disclosed
Sponsored content is labeled and every paid link carries rel="sponsored". What money buys here, and what it never does, is spelled out on the partners page and the rate card.
Where we make money, and the line we do not cross
To companies in the categories we publicly rank, we sell exactly two things: measurement (scans, reports, per-tool breakdowns of the numbers we already publish) and disclosed media (sponsorships and placements, every one labeled and marked rel="sponsored"). We do not sell them remediation. No sprints, no landing pages, no go-to-market work for a company that appears in our rankings.
Remediation services exist, but only for businesses outside the categories we rank, where we have no scoreboard to protect. The reason is plain: improving the AI visibility of a company we also rank would make us referee and coach at once, and the index's credibility is worth more than any engagement.
The constants underneath both sides: rankings and index positions are never for sale, sponsorship never changes a verdict, and every paid placement is labeled.
Do not take our word for any of this.
The monthly dataset is downloadable, the tracker state is a public JSON file, and every published figure traces to raw material we store. If you find a number this page cannot explain, email hello@claw.mobile and we will either explain it or fix it.