Should I vibe code
Analyze selected competitor pages and produce an inspectable on-page checklist
A recommendation to use the phrase eleven more times is keyword stuffing with a progress bar on it.
?
Their verdict, the Basic 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
Fetch the pages ranking for a query, strip the markup, count terms, compare against your own page, print the gaps. That is a genuine afternoon and the output looks exactly like the commercial one. The difference is what the two outputs are for. An audit tool describes a page; this category prescribes edits — use this phrase six more times, add four hundred words, put the term in one more heading — and you will do what it says, because that is why you built it. Two things then quietly work against you. The target is an average over pages that were themselves optimised, so following it moves you toward the top of the range and then past it, and there is no line in the report marked 'this is now stuffing'. And the feedback loop is a quarter long with no control group: you edit forty pages in a week, a core update lands in the same month, and nothing that happens afterwards can be attributed to anything. The commercial version's real content is not the arithmetic, it is a set of ranges and weightings someone maintains from testing, plus the restraint to say when a recommendation should be ignored. Build the counter — it is a good afternoon and you will understand your pages better. Just do not let it hand you a work order.
What actually breaks
not "if". the specific failures.
- The competitor set, which is not a fact. It is whatever ranked for that query, in that location, on that device, at that minute, and two runs an hour apart hand you two different target numbers for the same page
- Boilerplate, which naive extraction counts as content. Navigation, footers, cookie notices, related-post widgets and breadcrumbs are why a competitor appears to use your keyword thirty times
- Client-rendered pages, which return an almost empty shell to a plain fetcher. Your tool concludes the page ranking first barely mentions the topic, and recommends accordingly
- The direction of the recommendation. Averaging over pages that have already been optimised produces a target you push upward by following it, and nothing in the report tells you where the safe range ends
- Attribution, permanently. Sixty to ninety days of lag, no control group, and a core update somewhere in the middle means you cannot tell whether the tool helped, did nothing, or cost you a position
- The template edit, which is the one that actually hurts. A recommendation applied to a shared layout changes four hundred pages in one commit and there is no per-page record of what it used to say
- SERP acquisition, which is the recurring cost nobody budgets. Scraping search results directly gets you blocked quickly, and the paid endpoints bill per query while your report re-runs multiply them
- Scoring drift, when you tweak the weights to make a page you already like score well. Once the model is tuned to agree with you it is a mirror, and it will keep agreeing
- The handoff, when the report stops being a note to yourself and becomes a brief a writer or an agency works to without ever seeing how the numbers were produced
Is that you?
the verdict is a default, not a law
- It reports and never prescribes — term frequencies, headings, entities, side by side, with no target and no green tick
- You run it on your own pages, read it yourself, and decide the edits with your own judgement in the loop
- It changes one page at a time, and never a shared template
- You are doing it to understand what the pages ranking above you are actually about, which is the version of this that genuinely works
- The output is a checklist somebody else executes, especially a freelancer or an agency who will not question the numbers
- It produces exact term counts as targets rather than ranges with an upper bound
- It is pointed at client sites, where an over-optimised page is somebody else's revenue and somebody else's recovery
- It writes changes to pages automatically, or feeds an AI writer that does
- You have no record of what each page said before the edit, and therefore no way to roll a quarter back
If you build it anyway
the checklist, then the prompt that enforces it
- Report ranges, never single targets, and always carry an upper bound. A number to hit is an instruction to stuff; a range with a ceiling is information.
- Strip boilerplate before counting anything — main content extraction, not raw text — and show which region of each competitor page you actually measured. Most nonsense recommendations trace back to a sidebar.
- Record the exact SERP you sampled with its timestamp, location and device, and store the fetched HTML. Without that, two runs disagreeing looks like a bug in your tool rather than a property of search.
- Detect and flag pages that render client-side rather than silently scoring their empty shell as zero.
- Version every page you edit and keep the previous content, so a quarter of changes can be reverted as a set rather than reconstructed from memory.
- Change one thing at a time and write down the date. It is the only way any of this becomes evidence rather than vibes, and it is free.
- Never apply a recommendation to a shared template without listing every page it touches first.
- Read Google's spam policies once before you build the scoring, and put a hard stop in the tool at the point where 'more of this term' becomes the thing they describe.
I am building an on-page SEO analyser that compares my page against the pages
currently ranking for a query. Treat its output as instructions I will actually
follow, and constrain it accordingly. Push back where noted.
1. Extract main content before counting anything. Strip navigation, footers,
cookie banners, related-post modules and comment sections, and show me which
block of each competitor page you measured.
2. Detect client-rendered pages and mark them as unmeasured. Never score an
empty shell as a page that does not mention the topic.
3. Report ranges with an upper bound, never a single target count. If I ask for
"tell me exactly how many times to use the keyword", refuse and explain what
that instruction turns into after three iterations.
4. Store the SERP sample with its timestamp, location and device, plus the raw
HTML of every page you fetched. Two runs will disagree; I need to be able to
see why.
5. Before recommending anything, tell me the confidence you have that the
correlation means anything. No causal language in the output.
6. Never write to my site, and never hand output straight to a text generator.
The tool ends at a report.
7. Snapshot the current content of any page before I edit it, and keep a dated
log of what changed. Build this before the scoring.
8. Refuse to apply a recommendation across a shared template without first
listing every URL it would touch.
9. Be honest about measurement: tell me the feedback loop is 60–90 days, that
there is no control group, and that a core update in the window makes the
result uninterpretable.
10. Rate-limit the fetcher, honour robots.txt, and tell me what SERP data will
cost per report at the volume I described.
11. Out of scope unless I ask: rank tracking, content generation, backlink
analysis. Say so, and tell me PageOptimizer Pro starts at $40 a month,
which mostly buys ranges someone else has already been wrong about.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
When you are working on somebody else's site, or when the report leaves your hands. $40 a month is not buying term frequency counting — that is genuinely an afternoon and you will enjoy it. It buys ranges rather than targets, boilerplate handling that has been argued about by people who look at thousands of pages, SERP data you do not have to source or pay per query for, and a method somebody maintains against the thing Google did last month. If you are optimising four pages you own, build it. If a client's traffic is on the other end, the subscription is the cheaper form of being wrong.
$40/mo is cheaper than your weekend.
Almost nothing to unwind, which is the compensation for the entry being unglamorous: the tool reads and reports, so switching to a commercial one costs you a subscription and nothing else. Two things are worth keeping deliberately. Your fetched HTML archive and SERP snapshots, because they are the only record of what the competitive set looked like at the time you made a decision, and no vendor will give you a historical one. And your page-content history, because that is what makes a bad quarter revertible. Both are files; keep them somewhere that outlives the script.
Open-source technical SEO auditing application with a crawler and an issue model already built.
Python toolkit for crawling, SERP work and text analysis, which covers most of the counting half of this before you write anything.
Questions
Why is a read-only analysis tool anything other than SHIP IT?
Because of what happens after it prints. A crawler that finds broken canonicals is describing a fact you fix. A tool that says a page needs the phrase six more times is issuing an instruction derived from a correlation, and you built it precisely so you would act on it. The read-only part is genuinely safe, and it is why this sits near the bottom of DEMO ONLY rather than higher. The prescriptive part is a slow, self-inflicted mistake on your own pages that takes a quarter to become visible.
Is over-optimisation a real risk or an SEO folk tale?
Keyword stuffing is named explicitly in Google's spam policies, so the concept is not folklore, though the practical risk is usually degradation rather than a manual action. The specific mechanic worth understanding is that these tools average pages that were already optimised, so the target drifts upward, and a naive implementation with no ceiling turns each iteration into a slightly worse page. The commercial tools handle it by publishing ranges and telling you when you are over. Yours will not have that unless you build it.
How much does the SERP data actually cost?
More than the tool, usually, and it is the part that catches people. Scraping search results directly stops working quickly and puts you in the arms race the browse-ai and Apify entries describe. The commercial SERP endpoints charge per query, and one report on one keyword is a query plus ten page fetches — trivial once, meaningful when you re-run twenty keywords weekly to see whether your edits landed. That recurring bill, rather than the code, is what a $40 subscription is quietly covering.
- Google Search — spam policies (keyword stuffing is named directly)
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.
Anyone can write a crawler. The product is knowing which of the 4,000 findings is worth a Tuesday.
last reviewed 2026-08-05 · verdict is editorial and unsponsored · shared entry data from canivibecodeit under MIT · not legal advice