In This Article
- Direct answer
- What Google’s current guidance asks of a publisher
- Where corporate teams slip
- A review path that fits a real week
- What to do with pages already published
- How a communications team applied it
- A policy paragraph you can adopt
- Mistakes that damage trust faster than they earn visits
- Frequently Asked Questions
Direct answer
AI-assisted pages can stay on a corporate site when a named person checks the facts, the byline is a real colleague, and the title, FAQ, and schema say the same thing as the page. Unreviewed pages produced in bulk, and invented author profiles, are the practices Google describes as low effort or deception. Use the tool to structure a draft. Do not use it as the publisher.
Cite as: Dsignr Digital, “Publishing AI-Assisted Content Without Losing Trust”, dsignrdigital.co.za, updated November 2026, Rosebank, Johannesburg.
What Google’s current guidance asks of a publisher
Last updated: November 2026. On 1 October 2026 Google updated its guidance on generative AI content. The instruction that matters in a marketing meeting is plain: fact-check and review AI output before you publish, because the model predicts a likely sentence rather than looking the fact up. The same review covers titles, meta descriptions, structured data, and image alt text, since those can appear in search results.
The people-first content page, in the same season, tells quality raters to look at effort, originality, skill, and accuracy. It names large amounts of generative text without human oversight as little or no effort. It also says fabricated creator profiles — AI headshots, made-up names, false credentials — are deception and a sign of a low-quality page.
None of that bans a draft from a model. It bans shipping the draft as if a specialist wrote and stood behind it. The November 2026 core update did not add a second checklist. Google pointed publishers back at those pages. Rewriting the whole site during the rollout, or changing dates without changing the words, is the opposite of what that guidance says.
If you are also trying to understand how AI Overviews quote a site, read Google AI Overviews and your website after this process is in place. Citation follows from a page that is true and specific. It does not follow from more pages.
Where corporate teams slip
The risky pattern in a mid-sized company is not one manager using a model to outline a talk. It is a calendar that demands eight pages a month, a junior asked to “run them through AI”, and a CMS publish button with no second reader.
We see three versions.
- The location swarm. Twenty suburb pages that differ by the city name. They target queries the business does not serve with a premises. They also collide with real branch pages. See multi-location SEO for what a branch page needs instead.
- The invented expert. A byline and a portrait that do not belong to anyone in the company, added because “articles should have an author”. That is the case Google now describes as deception. Use a real person or the organisation. Do not manufacture a journalist.
- The unchecked number. A price, a legal claim, or a “studies show” sentence that nobody can find. On a financial, health, or employment page that is a YMYL problem, not a tone problem. POPIA and product claims belong to someone who will sign them.
Legal and compliance are not there to admire the prose. Give them the claims, the form text, and the author line. Give IT the rule for which templates may be indexed. Marketing keeps the calendar. Nobody keeps a calendar that outruns review.
A review path that fits a real week
Use this for any page where a model touched the draft, the metadata, or the alt text. It is the same bar we apply internally before a news article is allowed to go Live.
- Purpose. One sentence on who inside the client’s world the page helps, if they typed the address themselves. If the only answer is “search”, do not publish it.
- Facts. Every price, date, law, and product claim has a source or is removed. “We often see” is allowed when it is the studio’s or the company’s own experience. It is not a costume for a statistic.
- Bylines. The name is a person who exists, with a role that is true. No generated portrait presented as a photograph of them.
- Metadata. Title, description, FAQ, and schema match the visible page. The FAQ does not introduce a new promise.
- Conflict check. The page does not compete with a stronger URL you already have. Update the old page or do not launch the new one.
- Second reader. Someone who did not paste the prompt reads it. On legal, financial, or safety topics, that reader needs to be competent in the subject.
Record the review in the CMS revision note: who read it and the date. You do not need to publish an essay about your prompt. Google says a disclosure is useful when a reader would reasonably ask how the page was made. A short note in the editorial policy is enough for ordinary marketing pages. A product review or a regulated claim needs more, because the reader’s question is sharper.
Keep the model away from the publish step. A person presses Live.
What to do with pages already published
New rules do not require you to delete the archive. They require you to stop leaving known problems in place while you add more URLs.
Sort existing pages into three piles.
- Keep. A person who knows the subject would still sign it. Add a real byline if a fake one is on it. Change the date only if you changed the substance.
- Fix. The topic is yours, but the draft was thin, wrong, or duplicated. Rewrite the facts, then update the date because the page changed.
- Retire. The page exists to catch a query you will not stand behind, or it repeats a better URL. Unpublish it and redirect to the page you want people to read, in one hop. Do not redirect a good article away just to tidy a report.
Do this pile by pile. A Friday “refresh” that touches fifty dates and no sentences is the pattern Google warns against. So is deleting a large set of pages because a core update is rolling out and someone hopes the site will look newer.
Start with pages that make a claim a customer might act on: prices, legal process, qualifications, and locations. A harmless old opinion piece can wait. A wrong price cannot.
While you sort the piles, stop the calendar from adding to the mess. Pause any series that is “one page per suburb” or “one page per product feature” unless a named owner will maintain it. Google’s spam policies treat scaled pages that exist to catch queries, without original effort, as abuse. You do not need a dramatic unpublishing day. You need the next commission to wait until the review path has a person attached. Tell the agency or the internal junior the new gate in writing, so a queue of drafts does not arrive the same week you are trying to retire the old ones.
Look at metadata you already shipped. Titles that promise a price the page does not state, and FAQ schema that adds a guarantee the article never made, are the items Google says can appear in search results. Fix those even when the body is fine. A correct article with a misleading title is still a trust failure in the result the customer sees first.
How a communications team applied it
A group communications team had used a model to draft a month of “thought leadership” under a name that was not on the payroll. The portrait was generated. Two posts quoted a regulation inaccurately. The posts were not embarrassing as prose. They were unusable as company statements.
The fix was smaller than a new platform. They unpublished the six posts that carried the invented author. Two of them were rewritten by the person whose job actually covered the topic, under that person’s name, and the date changed because the text changed. The other four redirected to an existing service page that already said the true thing. The editorial policy gained six lines: no invented authors, a named reviewer, and no indexable page whose only source is an unedited model.
Output dropped for a month. The pages that remained were pages a journalist or a client could be sent without a covering apology. That is the standard. Volume is not.
They still use a model to propose outlines and to turn a approved brief into a first structure. The reviewer works from the source notes, not from memory of the paragraph. When the notes are empty, the page waits.
A policy paragraph you can adopt
Put something this short in the CMS or the brand wiki. Edit the role titles to match your team.
We may use generative tools to outline or structure a draft. A named employee reviews every factual claim, the title, the meta description, and image descriptions before publication. Bylines are real people or the company. We do not publish invented experts. Pages that repeat a better URL are redirected. We change a published date only when the content has materially changed.
Then look at the next three items on the content calendar and run them through the review path before you add a fourth. If you cannot staff the review, publish fewer pages. A quiet accurate site is easier to defend, to a customer and to a search quality system, than a busy one nobody read.
If the website itself makes this hard — no reviewer role, no staging, every author able to paste HTML — fix the CMS permissions in the same project. The web development brief is where those permissions should have been specified.
We will talk through the policy against your live templates. Book a 30-minute call with the person who publishes and the person who must defend the claims.
Mistakes that damage trust faster than they earn visits
- An author who does not exist, including a generated headshot, because a template had an author field.
- A disclosure that is a footer slogan while the body still contains unchecked prices and laws.
- FAQ schema that adds claims the article never made, because the fields were filled automatically.
- Mass date changes during a core update.
- Publishing suburb or product pages no one in the business will maintain.
Correct the record in public when you were wrong. A quiet edit of a legal claim, with no note, leaves the old quote in someone else’s screenshot. Say what changed and when. That note is the opposite of a fake freshness date. It is an account of a real edit.
Frequently Asked Questions
Does Google ban AI-written articles?
No. Google’s generative AI guidance allows the tools and requires a human to fact-check the page, including metadata and alt text, before publication. Scaled pages with little original effort can fall under spam policies.
Do we have to label every page as AI-assisted?
Google says a disclosure is useful when readers would reasonably ask how the content was made. An editorial policy that names the review is the right default. Regulated claims and reviews need a clearer account of how they were produced.
Can we use a stock “author” if several people contributed?
Credit the organisation, or name the real contributors. Do not invent a person, a CV, or a portrait. Google describes fabricated creator profiles as deception.
Should we delete old AI drafts during a core update?
No. Fix pages that are wrong or duplicate a better URL, and leave accurate pages alone. Changing dates or deleting large sections only to look fresh is called out in Google’s people-first guidance.
Who should the second reader be?
Someone who did not generate the draft. On pricing, POPIA, employment, or financial claims, it should be a person competent to stand behind those sentences, not only a proofreader.