Is ChatGPT good at SEO? Where general AI falls short

Ask ChatGPT for a keyword strategy and it performs beautifully. Within seconds you have a tidy table: keywords, monthly search volumes, difficulty scores, content suggestions. It looks exactly like the output of a $100-a-month SEO tool, and it cost you nothing.

The problem is the numbers. ChatGPT has no database of search volumes. It has never seen your analytics. It cannot check today's rankings. The volumes in that table were generated the same way it generates sentences: by producing plausible-looking text. Ask again tomorrow and you'll often get different numbers for the same keywords, because there was never a lookup happening, only prediction.

So is ChatGPT good at SEO? The real answer is a split decision, and knowing exactly where the split falls is worth more than either a yes or a no.

What ChatGPT is genuinely good at

Start with credit where it's due, because the good half of the split is very good.

Drafting. SEO runs on written pages, and producing them is the single most expensive part of doing SEO yourself. This is where language models measurably shine: in a controlled MIT experiment published in Science, 453 professionals given ChatGPT finished writing tasks about eleven minutes faster with output judged 18% better by independent evaluators. Today's top-ranking pages reflect this: Ahrefs found 82.2% of top-three results contain AI content blended under a human hand. Used with a proper brief, ChatGPT collapses the cost of the most expensive input in SEO.

Explaining. Ask what a canonical tag does, why a 301 redirect matters, or what "search intent" means, and you'll get a clear, patient answer at whatever level you need. As a tutor for the vocabulary of SEO, it's excellent, because explanation is a text-pattern job and settled concepts don't go stale.

Brainstorming raw material. "List forty questions a homeowner might ask about heat pumps" produces a genuinely useful pile of topic candidates in seconds. The pile is a starting point, not a plan: which of those questions people actually search for, and how often, is a data question, and that's the other half of the split. The reliable workflow is ChatGPT for breadth, then real search data to filter it.

Grunt work. Title tag variants, meta descriptions, FAQ phrasings, rewriting a paragraph to answer a question directly, drafting schema markup for a human to validate. Low-stakes, text-shaped, endless. Perfect delegation material.

Where it falls short, structurally

The failures aren't bugs that the next version will fix. They follow from what a general chatbot is: a text predictor with no data connections and a training cutoff.

It has no search data. Search volumes, keyword difficulty, click rates, trends: these live in databases owned by Google and the SEO tool vendors. ChatGPT contains none of them. When you ask for volumes, it produces numbers shaped like the ones in its training text. They're not estimates. They're fiction with formatting. Every model has a documented training cutoff date, and even within its training window it stored patterns, not tables.

It can't see your site. It has never read your Search Console, your analytics, or your server logs. It doesn't know which of your pages are indexed, which queries you already rank for, or which posts bring in your customers. The free tool that knows all of this is Google Search Console, and the difference matters: ChatGPT gives advice for a generic site, while your data gives advice for yours.

It can't see today's results. SEO happens against a live, shifting scoreboard. Base-model ChatGPT sees nothing after its cutoff: no current rankings, no algorithm update from last month, no new competitor. The browsing-enabled versions can fetch pages and help here, but a handful of fetched pages is still spot-checking, not the systematic tracking that tells you whether things are improving.

It fabricates specifics. This is the dangerous one, because the failures are invisible. When models are asked for sources and statistics, the measured invention rates are high: a 2024 medical study found GPT-4 fabricated 28.6% of the references it produced, and a 2026 audit of nearly 70,000 citations across ten models found hallucination rates from 11.4% to 56.8%. SEO advice comes wrapped in exactly this kind of specific: "studies show", "Google confirmed", "the average CTR is". OpenAI's own interface says it plainly: "ChatGPT can make mistakes. Check important info."

Its advice is the average of the internet. ChatGPT's SEO recommendations are a summary of what SEO blogs say, including the outdated and wrong parts, delivered identically to you and your competitor. Sometimes that average is fine. But Google's own 2026 guidance had to explicitly warn against fashionable hacks, stating that "you don't need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search", after a wave of AI-recommended busywork like llms.txt files, which Google says it ignores. A chatbot trained on the internet's SEO chatter faithfully reproduces the chatter, hacks included.

The tell: confidence without a source

Here's a practical detector for the bad half of the split. When ChatGPT's answer contains a number, a study, or a "Google says", ask where it came from. A data tool can tell you: its own index, its own crawl, Google's API. ChatGPT will either admit it can't verify, or, worse, generate a citation, and generated citations are fabricated at the rates above.

The rule that follows: use ChatGPT's words, never its numbers. Any figure that matters gets confirmed in a primary source or a real tool before you act on it, the same discipline as fact-checking an AI-written article.

Three experiments you can run in ten minutes

Don't take the argument on trust; the limits are demonstrable on your own account, and seeing them once inoculates better than any article.

The consistency test. Ask for monthly search volumes for five keywords in your industry. Open a fresh chat and ask the identical question. Compare. Real data sources return the same number twice; a text predictor returns numbers shaped like answers, and the drift between the two chats is the fabrication made visible. (If the tool searched the web and cited a keyword-data source, note what actually happened: the data came from a real tool, retrieved, which is precisely the division of labor this article recommends.)

The source test. Take any statistic ChatGPT has given you and reply: "Link me to the primary source for that figure." Then click what it produces and read the page. Sometimes the link is real and says something adjacent; often the page doesn't exist or doesn't contain the number. Either way you've watched the citation machinery at work, and you'll never again paste a model's statistic without checking.

The mirror test. Ask it something about your own site that you know cold: "What does [your business] rank for?" or "How much organic traffic does [your site] get?" The answer will be fluent, plausible, and, compared against your Search Console, wrong in ways only you can see, because you hold the real data. Now remember that its answers about everything else are generated by the same process. The only difference was that this time you could check.

What about the newer, search-enabled versions?

A fair objection: ChatGPT can now browse, and models keep improving, so isn't this article aging badly? Partly, and it's worth being precise about which limits are version problems and which are structural.

Browsing genuinely fixes staleness for the queries where it triggers: a search-enabled assistant can fetch today's page, cite it, and be usefully current. What it doesn't create is a dataset. Fetching five pages about keyword difficulty is not the same as owning an index of query volumes, rankings, and click curves, and no amount of model improvement turns a text interface into your Search Console, your analytics, or a crawl of your site. Fabrication, meanwhile, is being reduced version by version but remains a documented property of the technology, which is why the vendors' own disclaimers persist. The stable rule across every version so far: trust the words more each year, verify the numbers every year.

The division of labor that works

None of this argues for abandoning the tool. It argues for a specific pairing, and the pairing is cheap.

• Topic ideas: ChatGPT for the brainstorm, then Google Search Console and a keyword tool to find which ideas have real demand.
• Writing: ChatGPT drafts from your brief, you edit and verify. Its best SEO job by far.
• What's working: Search Console, free, tells you your real queries, impressions, and clicks. ChatGPT can help you interpret an exported report, which is a text job, once the numbers come from somewhere real.
• Rankings and competitors: a tracking tool or manual checks. Not ChatGPT, which will cheerfully describe a SERP it cannot see.
• Technical checks: crawlers and validators for the facts. ChatGPT to explain what the findings mean and draft the fix.
• Strategy: your data plus your judgment. ChatGPT as a sounding board that knows the textbook, not as the decider, because it doesn't know your market, your margins, or your customers.

The pattern in every row: ChatGPT handles language, real tools handle measurement, and you handle decisions. Where businesses get burned is letting the language machine quietly take over the measurement column, because it never refuses the job and never mentions it's guessing. This division is also what a sensible budget SEO setup looks like: most of the paid-tool spend goes to data, because words got cheap.

So: is it good at SEO?

At the writing inside SEO, yes, genuinely, and that's not faint praise, because writing is most of the work and most of the cost. At the data inside SEO, no, and it fails in the worst possible way: fluently, confidently, and without a warning label beyond the small print.

Treat ChatGPT as the most productive writing assistant you've ever had, wired to a compulsive tendency to invent statistics. Pair it with Search Console for truth about your site and a real keyword tool for truth about demand, keep every number on a verify-before-use rule, and you'll get the whole upside without ever betting your strategy on a hallucinated table.