Should I vibe code
Search and AI visibility platform with TrueDR, monitoring, and verified partnerships
Ask a model the same question twice and get two answers. Chart the difference and you have invented a trend.
?
Their verdict, the Pro price and the build-time estimate come from their entry, MIT-licensed. Checked 2026-08-04.
?
Our verdict, the regret score and everything below it. Editorial and unsponsored — nobody can pay to be moved.
The honest answer
why the verdict is what it is
Despite the name, there is no doctor in this one — DR is Domain Rating. VerifiedDR asks ChatGPT, Perplexity and Google's AI mode a list of buyer questions each week, records which brands get recommended, verifies claimed backlinks by refetching them, and scores the result. Two thirds of that is a cron job an agent will write in an evening, and it will look convincing. The problem is that the chart is mostly noise: the same prompt to the same model on the same afternoon returns different brands, so a five-point drop in your "share of voice" is usually sampling variance, and what you have built is an instrument that manufactures decisions out of it. The other third is worse. Reading Google's AI mode means scraping Google, which is a blocked-proxies arms race that breaks on a schedule you do not set, and the partnerships half — matching you with sites to place links on — is link building, which is a thing Google's spam policies address by name. Build it to satisfy your curiosity about one brand. Do not build it to hand somebody a number.
What actually breaks
not "if". the specific failures.
- Determinism, because the same prompt to the same model returns different brands within the same hour, and your week-on-week delta is largely that
- Scraping Google's AI mode, which is a blocked-IP problem wearing a data problem's clothes and which stops working on Google's schedule
- Your model bill, when 25 prompts across three platforms quietly becomes 200 across six and nobody put a ceiling on the job
- Brand matching, which counts a competitor's passing mention, a Reddit quote and a hallucinated near-miss of your name as the same event
- Backlink verification, which means politely fetching thousands of other people's pages — and doing it impolitely is how you end up blocked or reported
- The composite score you invent, which somebody will eventually ask you to defend line by line, and which is a weighted guess
- Silent failure, where a parser change turns every answer into "not mentioned" and the dashboard renders that as a catastrophic collapse
- The partnerships feature, which is a link exchange with a nicer name, and which is squarely inside Google's link-spam policy
Is that you?
the verdict is a default, not a law
- It watches one brand, for you, and you read it as a rough signal rather than a metric
- Every data point stores the raw model response next to the verdict, so you can see why the line moved
- You sample each prompt enough times to have an honest error bar, and you draw it
- It only calls APIs you are permitted to call, and never scrapes a results page
- You are selling the number to clients, because the first question will be what a three-point drop means and there is no answer
- It scrapes Google, which will work right up until it does not and will not tell you which day that was
- The output is a list of sites to buy links from, which is a different business with a different risk profile
- There is no cost ceiling on a weekly job whose cost is prompts times platforms times samples
- You will make budget decisions from it without ever looking at the raw responses underneath
If you build it anyway
the checklist, then the prompt that enforces it
- Sample every prompt several times per run and store all of the responses. One sample is an anecdote; the variance between them is the actual finding.
- Show uncertainty on the chart. A line without an error band on non-deterministic data is a graph that lies politely.
- Keep the raw model output forever, alongside the parsed verdict. When the number moves, the only useful debugging artefact is what the model actually said.
- Put a hard spend cap and a run counter on the job before the first scheduled execution, not after the first invoice.
- Use official APIs. If a source has no API, treat that as the source saying no rather than as a scraping exercise.
- Crawl other people's sites politely: respect robots.txt, set a real user agent with contact details, rate limit per host, and cache aggressively.
- Alert on a suspicious zero. Every prompt returning "not mentioned" is far more likely to be your parser than your marketing.
I am building an AI-search visibility tracker: ask a set of buyer questions to
several models each week, record whether my brand appears, and chart it. Build
it so the chart cannot lie to me, and argue when I ask for the pretty version.
1. Start by telling me plainly that these models are non-deterministic, that
identical prompts return different brands, and that most week-on-week
movement in a single-sample design is noise rather than signal.
2. Therefore: multiple samples per prompt per run, all responses stored, and
every chart drawn with a confidence interval. No single-sample line graphs.
3. Persist the raw model output next to the parsed result, always. When a
number moves, the raw text is the only thing that explains it.
4. Put a hard spend cap and a maximum-calls-per-run limit in before you
schedule anything. Show me the projected monthly cost of my configuration.
5. Use official APIs only. If I ask you to scrape Google's AI mode or a results
page, refuse and explain that it breaks silently and gets addresses blocked.
6. Any crawling of third-party sites obeys robots.txt, sets a user agent with
my contact details, rate limits per host, and caches. I am a guest.
7. Brand matching is explicit and reviewable: exact name, known variants,
domain mentions, each counted separately. Never collapse them into one
number without showing the breakdown.
8. Add a sanity alarm: if every prompt in a run returns "not mentioned", assume
my parser broke before assuming my visibility collapsed.
9. Do not invent a composite score. If I insist, make its formula visible on
the page next to the number and refuse to weight anything I cannot explain.
10. Out of scope: link-buying, outreach automation, and anything that places
content on other people's sites. Tell me that is a separate decision with
Google's spam policies attached.That one keeps you out of trouble. For the prompt that actually builds it, canivibecodeit.com has one.
their build prompt ↗Or don’t build it
the boring option, and the way back out
If you need the number to be defensible to somebody else, or you want the Google-side data without operating the scraping yourself. $39 a month buys the maintenance of an integration surface that changes without notice, and the crawl infrastructure for backlink verification, which is the part that turns from a script into a job. It does not buy determinism — nobody sells that — so read whatever you buy with the same suspicion.
$39/mo is cheaper than your weekend.
The valuable artefact is the archive of raw responses, not the dashboard. Store one JSON file per prompt per run — the model, the timestamp, the full text, the parsed verdict — in flat storage you could hand to anything else, and the visualisation becomes disposable. Anyone taking over gets a dataset they can re-derive their own metric from, which is worth considerably more than a chart they have to trust.
Self-hosted search-rank tracker with keyword monitoring and email reports; the classic-SEO half of this, already built and MIT licensed.
Local-first open-source dashboard that tracks brand visibility across several AI answer engines with mentions, position and citations.
MIT-licensed AI visibility tracker that drives real browser sessions against ChatGPT, Gemini, Perplexity and AI Overviews rather than APIs.
Questions
The DR in the name — is this a medical thing?
No. DR is Domain Rating, the backlink-authority metric popularised by Ahrefs, and the product's pitch is a "true" version of it that discounts links it cannot verify by refetching the page. No clinicians, no patient data, nothing regulated. If you arrived expecting doctor verification, this is the wrong entry.
How much variance are we actually talking about?
Enough that a single sample tells you very little. Answer engines re-rank sources between requests, personalise on context you cannot see, and change model versions without announcement. If you run the same ten prompts twice in an afternoon and count brand appearances, the two runs will disagree — so the honest design samples repeatedly and reports a range, and any product showing you a smooth weekly line from one sample per prompt is smoothing over the thing you wanted to measure.
Is the link-partnership part really a problem?
It is at least a different decision than the tracking part, and worth making deliberately. Arranging placements on other sites for the purpose of moving rankings is exactly what Google's link-spam guidance describes, and the downside is not a fine but a manual action against your own domain — which is slow to notice and slower to unwind. Build the measurement half; think much harder before automating the acting-on-it half.
Every week, someone ships something they shouldn’t have.
New verdicts, the worst thing that landed in the trap, and the occasional incident report. No other email, ever.
SERP analysis means scraping search results, which is a fragile foundation to build on.
Crawling your own site and reading the headers is a script, not a licence.
Anything that mines a platform for ranking signals is renting its data supply from someone hostile.
last reviewed 2026-08-04 · verdict is editorial and unsponsored · shared entry data from canivibecodeit under MIT · not legal advice