Myth vs reality

Your assistant quoted a backlink count. Where did it come from?

Usually from nowhere. A model with no tool attached writes the most likely number, and a likely number is not a measured one.

SEO AI Pal research team · Updated 2026-10-29

The blunt version

No tool call, no measurement

A chat window on its own cannot see the web's link graph. It predicts text. If you ask how to check number of backlinks in a website and it answers with a figure for your domain, treat the figure as a sentence, not a reading. Fixing this takes one real call.

Six things people believe about assistants and link data

Left column is what we hear. Right column is what actually happens.
Myth	Reality	What to do
"It browsed my site, so the count is live."	Browsing a page shows the page. Links pointing at it live on other sites, and a page fetch never lists them.	Ask for a lookup through a link index, not a page read.
"It gave the same number twice, so it is right."	Models repeat themselves. A stable wrong answer is still wrong.	Compare against a call you can reproduce with curl.
"The number came with a confident tone."	Tone is style. Confidence says nothing about whether any lookup ran.	Require the reply to quote the `_credits` field of the response.
"A round number means it estimated honestly."	Round numbers are what models produce when guessing. Real counts are untidy.	Distrust 1,000, 5,000 and 10,000.
"Free chat plus Search Console equals a link report."	Search Console lists links Google chooses to show you. It is useful, and it is a sample.	Use it as a cross-check, not the only source.
"One tool call covers backlinks, anchors and rating."	Each is a separate action here: `refdomains`, `anchors` and `overview` are different calls.	List the actions your assistant may call, one per job.

What a real reading looks like

Real data arrives as JSON with an ok flag. Authenticated calls also carry _credits, which shows the cost and your remaining balance, and a model cannot invent that field convincingly because you can compare it with your dashboard.

Try it once by hand. Run curl "https://seoaipal.com/api.php?action=overview&q=example.com&key=sk_live_..." in a terminal. It costs nothing. Now paste the same domain into a bare chat and ask for the count, and watch the two disagree.

Cadence matters. The index behind it is our own: 181,132,097 domains and roughly 798.5 billion backlinks, refreshed about once a month. That also means a link won last week may be missing. A fresh link absent from the reply is a refresh gap, not a hallucination, and the two failures look different once you know the cadence.

Where assistants do help

They are good at the dull parts around a lookup: turning a domain list into calls, reading a table of rows, writing the sentence for a client. Give them the numbers. Do not ask them to remember numbers. The setup is in the custom GPT walkthrough.

Five quick tests that expose a guessed count

Ask the same question in a fresh chat. A real lookup repeats exactly; a guess drifts by a few percent.
Ask which action produced the number. A guesser answers vaguely or names a product it never called.
Ask for the _credits value. No call, no balance.
Check the date. Our index refreshes monthly, so a reply claiming hourly freshness is wrong.
Run the free overview action yourself and compare the first figure.

The Ahrefs and Semrush question

People ask whether an assistant plus an API replaces Ahrefs or Semrush. It does not. Those tools crawl daily and cover keywords and rank tracking, which a link-only API does not.

Different job. What it replaces is the habit of pasting a screenshot of a dashboard into a chat and hoping the model reads it right. If your questions are about links, a direct call is cheaper and checkable, and from $0.10 per 1,000 calls the experiment barely registers on a card statement. For the failure patterns that follow, see the tool-call error table.