If you draft with AI, you now need a human pass before you hit publish — not a vibe check, an actual one that verifies every claim, number, name and link. Google's updated helpful-content guidance makes this plain: AI assistance is fine, unchecked AI output is not. The fix isn't hiring an editor. It's a repeatable 15-minute routine you run on every post.
The good news for a one-person team: most AI mistakes cluster in predictable places. Stats, quotes, dates, product names, and the bits you never look at — titles, alt text, meta descriptions. Learn where the model lies and you can check fast.
Below is the workflow I use, built for someone who ships a post a week and can't afford a fact-check desk.
What Google actually changed
Google didn't ban AI content — it clarified that AI output still has to meet the same bar for accuracy and usefulness as anything you'd write yourself, and that the responsibility to verify sits with the publisher. The key shift is scope: fact-checking now explicitly covers the parts people skim past, including titles, alt text and metadata.
In practice this means a confidently wrong statistic in your opening paragraph, a made-up study in a subheading, or a misleading title the model wrote to sound punchy can all drag a page down. It's less "did a human touch this" and more "did a human confirm it's true." Thin, unverified AI pages are exactly what the helpful-content signals are tuned to catch.
The practical takeaway: treat your AI draft as a fast first draft from a smart intern who occasionally invents things with total confidence. Your job is the review, and the review is non-negotiable.
Where AI invents things — and how to catch it fast
AI fabricates most often in five spots: statistics, quotes, named studies or sources, specific dates, and links. Check those five and you catch the large majority of errors in minutes.
Models produce text that sounds sourced because it mimics the shape of sourced writing. "A 2023 study found 68% of…" reads authoritative whether or not that study exists. So verify the shape, not the confidence. Here's the triage I run, fastest to slowest:
- Every number gets a source. Copy the exact stat into a search engine in quotes. If you can't find a primary source (the original report, not a blog quoting a blog), cut the number or replace it with one you can stand behind.
- Every quote gets attributed to a real, findable person. Search the quote verbatim. AI loves to attribute plausible-sounding lines to real experts who never said them.
- Every "study" or "report" gets a title and a link. If the model names "a Stanford study" with no title, assume it doesn't exist until you find it.
- Every date and version number gets checked — product releases, law dates, pricing. These drift and models train on stale data.
- Every link gets clicked. AI hallucinates URLs that look real and 404. Open each one.
Two free tools worth your time: Google Scholar for verifying any academic-sounding claim, and the original source's own site search for stats. If a figure only appears on content-mill sites and never on a primary source, it's probably a telephone-game distortion — drop it.
The five-minute stat audit
Pull every number out of the draft into a list, verify each against a primary source, then paste the verified figure back with the source linked. Doing it as a batch is faster than checking in context because you stay in "verify" mode instead of ping-ponging between reading and researching.
Open a scratch note and list each claim on its own line: the number, what it's about, and a blank for the source URL. Work down the list. For each one, find the primary source, confirm the figure matches, and paste the URL. Anything you can't source in about 60 seconds gets flagged for cutting — not kept "because it's probably right."
A quick rule that saves embarrassment: round numbers and suspiciously tidy percentages ("exactly 50% of businesses") deserve extra scrutiny. Real research is messy. If the stat is too clean, it's often invented or badly paraphrased.
Don't skip titles, alt text and metadata
These are the fields that now count — and the ones AI gets wrong most quietly, because you rarely read them after generating. Check that your title matches the actual content, your meta description doesn't promise something the page doesn't deliver, and your alt text describes the real image rather than a generic guess.
The classic failure: you ask AI for a title, it produces something with a number or claim ("7 Ways to Double Your Traffic") that your article never backs up. That's a mismatch Google can detect, and it erodes trust the moment a reader clicks. Titles must be honest to the body.
Alt text is the sneakiest. If you let AI write alt text without showing it the image, it describes an imagined picture. Bad for accessibility, bad for image search, and flatly inaccurate. Always write alt text against the image that actually ships.
| Field | Common AI error | 30-second check |
|---|---|---|
| Title | Claim the body doesn't support | Does the article deliver this exact promise? |
| Meta description | Overpromises or repeats keywords | Read it as a reader — accurate and inviting? |
| Alt text | Describes an imagined image | Open the image, confirm the description matches |
| Headings | Invented stat in an H2 | Treat every subhead claim like body copy |
A pre-publish checklist you can run in 15 minutes
Run this in order every single time, and fact-checking stops being a dreaded chore and becomes muscle memory. Keep it pinned next to your editor.
- Stats: every number has a primary-source link, or it's gone.
- Quotes: verbatim search confirms who said it.
- Sources: every named study has a real, clickable title.
- Links: clicked, loading, pointing where you meant.
- Dates and versions: current and correct.
- Title: honest to the content, no unsupported claim.
- Meta description: accurate, under ~155 characters.
- Alt text: written against the real image.
- Your own experience: at least one thing in here only you could have written.
That last line matters most. The thing AI can't fake is your hands-on knowledge — the mistake you made, the setting that fixed it, the client who taught you the hard way. Add one genuine detail and the whole piece reads human, because it is.
If you're publishing on a site hosted with us, you've got a small safety net for the technical side of the same discipline. Our AI chat is on any hour to answer "is this link broken on my end or yours," with real engineers following up in business hours — so you can tell a content error from a hosting one without losing your evening. And because TPC Hosting is EU-based and GDPR-friendly, you don't have to second-guess where your drafts and data live while you work.
Build it into your routine, not your willpower
The teams that keep this up don't rely on discipline — they bake the check into a template so skipping a step takes more effort than doing it. Make a reusable publish checklist in whatever you draft in, and refuse to hit publish until every box is ticked.
One more habit: keep a running "sources" note per post with every link you verified. If a claim is ever questioned, you've got the receipt in seconds. It also makes updating the post a year later far less painful, because you know exactly where each number came from.
AI makes the first draft cheap. Your verification is what makes it worth publishing. That gap — between generated and checked — is now the whole game, and it's one a careful solo operator can win without hiring anyone.
FAQ
Does Google penalise AI-written content?
No — Google penalises unhelpful, inaccurate content regardless of how it's made. AI assistance is fine, but the publisher is responsible for verifying every claim, and unchecked AI output that's thin or wrong is exactly what the helpful-content signals target.
How do I spot a fabricated statistic in an AI draft?
Search the exact figure in quotes and trace it to a primary source. If the number only appears on blogs quoting other blogs, or you can't find the original report, treat it as invented and either cut it or replace it with a stat you can link to directly.
Do I really need to fact-check titles and alt text?
Yes — Google's updated guidance explicitly includes titles, alt text and metadata. These are where AI quietly goes wrong, especially alt text written without the model seeing the real image, so give each a quick accuracy check before publishing.
How long should a pre-publish fact-check take?
About 15 minutes for a typical post once you batch it. Pull every number, quote, source and link into a list, verify them together, then run a quick pass over the title, meta description and alt text.
Can I automate fact-checking entirely?
No — automation can flag broken links and surface sources, but confirming a claim is true still needs human judgement. Use tools to speed up the search, then make the call yourself; that final verification is the part Google now expects from you.
