Should I vibe code
Research a topic, build a content brief, and draft against selected SERP concepts
Your tracker will report a 30% drop the week OpenAI changed its router. Nothing about your brand moved.
?
Their verdict, the Starter 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
Check the product before you copy it, because Scalenut is not the thing the tagline describes any more. The SERP-brief-and-draft workflow is still in there, but the company now sells itself as a GEO platform: it tracks whether your brand gets mentioned and cited inside ChatGPT, Perplexity, Gemini, Claude and Grok, and it staples a backlink marketplace and a bench of human strategists onto that. Two of those three are not software. The one that is — the visibility tracker — looks like the easiest thing in the world to build, and it is: fire a list of prompts at five APIs on a schedule, count the brand mentions, draw a line chart. The problem is what the line means. Those answers are sampled, not computed. They shift with temperature, with account, with country, with whichever model the vendor happened to route you to that morning, and none of the engines that matter most give you a stable, purchasable interface to the surface you actually care about. You will build a dashboard in a weekend, and then you will spend a year making content decisions off a metric that moves when nothing about your brand did.
What actually breaks
not "if". the specific failures.
- The measurement, first and worst: LLM answers are sampled rather than computed, so the same prompt on the same day gives different brand mentions and your chart records that as a trend
- The absence of a control group — there is no held-out version of your brand that did not publish the article, so nothing in the system can attribute a movement to anything you did
- Google AI Overviews, which have no public API, so the only way to track them is scraping the SERP, which is the exact behaviour that gets an IP blocked and pushes you onto a paid proxy
- Prompt drift: to compare week to week your prompt set has to be frozen, and the moment you improve a prompt the history before it becomes incomparable
- Personalisation and geography, which mean your tracker measures what your datacentre in Virginia sees, not what your customer in Munich sees
- The cost curve, once 200 prompts across 5 engines every week becomes 200 prompts across 5 engines every day because weekly felt too coarse
- Model deprecation, which retires the exact model version your baseline was measured on and resets your entire history
- The drafting half, at volume, where a pipeline that publishes a hundred near-identical pages to your own domain is the thing search engines now specifically demote
Is that you?
the verdict is a default, not a law
- You want a research and brief tool for yourself: pull the top results, extract the headings and entities, and produce an outline a human writes from
- The tracker is a curiosity you look at monthly, and nobody's budget moves because of it
- You report it as a range with a sample count, not as a number with a decimal place
- Every draft passes a human who is accountable for the claims in it
- The chart is going into a client report, because you are then selling a measurement you cannot reproduce
- You are publishing generated pages at volume against your own domain and treating the tracker as proof it worked
- The pipeline scrapes AI Overviews or SERPs at any volume and you have not read the terms you are agreeing to
- Anyone is going to be paid, promoted or fired against the number
If you build it anyway
the checklist, then the prompt that enforces it
- Sample properly or do not sample at all. Run every prompt n times, record the mention rate with its variance, and put error bars on the chart. A single run per prompt per week is a coin toss with a line drawn through it.
- Freeze the prompt set and version it in the repo. Any change to a prompt starts a new series; never silently backfill the old one.
- Record the model identifier, the date, the region and the temperature with every single observation. Without those columns your history is uninterpretable the first time a vendor swaps a model.
- Log the raw answer text alongside the mention count. When the number moves you will want to read what actually changed, and a count cannot tell you.
- Buy the SERP data rather than scraping it, or accept that you have built a scraper and will be maintaining it forever. Price both before you write either.
- Put a hard monthly spend cap on the API keys, in the provider console rather than in your own code. A tracking loop with a bug is a billing incident.
- Keep a human between the generator and publish. Scaled, near-duplicate output on your own domain is a risk to the domain, not just to the article.
- Write down, before you start, what number would change your behaviour. If no plausible reading of the dashboard changes what you do next week, you are building a screensaver.
I want to build an AI-search visibility tracker and content brief tool. My
biggest risk is not a bug, it is believing a number that is mostly noise.
Build for measurement honesty first and features second.
1. Before any UI, design the sampling: every prompt runs n times per
observation, and the stored record is a rate with a variance and a sample
count, never a single yes/no.
2. Every observation row carries the engine, the exact model identifier, the
timestamp, the region and the temperature. Refuse rows missing any of those.
3. Store the full raw answer text beside the extracted mention count, so a
movement can be read rather than guessed at.
4. The prompt set lives in version control. Changing a prompt starts a new
series; never backfill or silently merge histories.
5. Charts show error bars and sample counts by default. If I ask for a clean
single line, tell me why that line will mislead me.
6. Before adding Google AI Overviews, stop and explain that there is no public
API, that the alternative is scraping, and what that costs me in terms of
service and IP blocking. Make me choose deliberately.
7. Put a hard monthly spend cap on every API key at the provider rather than in
application code, and show me the projected cost for my prompt set at my
chosen cadence before the first scheduled run.
8. Rate-limit and back off on every engine. Retries on a schedule are how a
tracker turns into an accidental load test.
9. For the content half: produce briefs and outlines, not finished pages, and
mark every generated draft unpublished pending a human action.
10. Refuse to build a bulk publish button. If I ask, explain scaled content
abuse and what it does to a domain rather than to a page.
11. Put in the README what this tool cannot tell me: causation, personalised
results, and anything about users routed to a model I did not sample.
12. Out of scope: backlink acquisition and outreach. Scalenut sells those with
humans attached, from $59 a month, and that is the part you cannot code.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
Buy when the output is going in front of someone else. A client will not accept "my script counted mentions" as evidence, and $59 a month buys a methodology somebody else has to defend, plus the two things you genuinely cannot build: a backlink marketplace and human strategists. Build your own if it is your own site, your own curiosity and your own budget — a research-and-brief tool for one person is a good weekend and the tracker is a fun toy, as long as you never let it into a slide deck.
$59/mo is cheaper than your weekend.
Keep the observations in a plain table you own — one row per prompt, per engine, per run, with the raw text — and the history survives you abandoning the dashboard, which is the only part with any value. Briefs and outlines should be Markdown files in a repo rather than rows in an app, so the writing outlives the tool. The thing that does not survive is any comparison across a vendor change: when you switch tracker, your old numbers and the new ones are not the same measurement, and pretending otherwise is worse than starting the series again.
Self-hostable metasearch engine — the least painful way to get search results into a pipeline without buying a SERP API.
Open-source front end for local and hosted models with retrieval built in; covers the drafting half.
Questions
Has Scalenut actually changed, or is this just marketing language?
It has changed enough to matter. The current site leads with generative engine optimisation, tracks brand mentions across ChatGPT, Perplexity, Gemini, Claude and Grok, and sells a backlinks marketplace and a team of strategists alongside the software. The pricing page has been rebuilt around "prompts analyzed" as a quota. If you are copying the product described in an older review, you are copying something the vendor has moved on from.
How is this different from the NeuronWriter and Frase entries?
NeuronWriter's argument is unit economics — you cannot buy the SERP API input for less than the finished product costs. Frase's is that you can build a content score but not the authority that makes people obey it. This one is about measurement: the metric Scalenut now sells is sampled from non-deterministic systems, and a homemade version will present sampling noise as a trend with no way to tell the difference.
Can I not just call the APIs and count mentions?
You can, and it takes an afternoon. The trap is that it works immediately and looks authoritative. Run the identical prompt ten times against one model and count how often your brand appears — the spread you get is the floor on what your dashboard can detect. Anything smaller than that spread is not a signal, and most week-on-week movements are smaller than that spread.
What is the safest useful thing to build here?
A brief generator for one person. Pull the top ranking pages for a query, extract headings, entities and questions, and produce an outline you write from. It has no measurement claims to defend, it costs pennies, and it is the part of the original Scalenut that was genuinely worth the money before the product moved.
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.
The cheapest SERP API is $25 a month. NeuronWriter is $23. You cannot buy the ingredients for less than the meal.
You can build the scoreboard. You cannot build the referee — and the score is the thing people obey.
SERP analysis means scraping search results, which is a fragile foundation to build on.
last reviewed 2026-08-05 · verdict is editorial and unsponsored · shared entry data from canivibecodeit under MIT · not legal advice