August Spam Update and AI Content: How to Recover Without Losing AI Visibility

Google's published spam policies name two behaviours that matter here: scaled content abuse, meaning pages produced in volume primarily to manipulate rankings rather than to help people, and site reputation abuse, meaning third-party content hosted on a domain to borrow its authority. Those are the policies to quote in an internal report, because they are specific and testable, whereas the word "quality"invites weeks of unproductive debate. The Google Search spam policies do not prohibit AI assistance in drafting, and Google has repeatedly said as much. The distinction matters because it changes what a remediation plan should contain.

The affected sites share recognisable patterns. Templated page families where only a city name, product attribute or job title changes. A publishing velocity that jumped from a handful of posts a month to dozens a week without any corresponding increase in editorial capacity. Near-identical opening paragraphs across a cluster. No named author, or a stock author profile with no verifiable background. No first-hand evidence anywhere on the page: no original figures, no tested findings, no dated observation, nothing a reader could not have obtained from the top three results already ranking.

"We use AI responsibly"is not a defence, because it describes an intention rather than an artefact. The practical test is applied page by page: does this URL contain something that could only have come from this business. Publishers running AI-assisted workflows with genuine human review, subject-matter input and original data reported little or no movement, and that is the clearest available signal about what was actually being demoted. Human review before publishing is not a compliance ritual. It is the step that adds the evidence the classifier is looking for.

Grounded AI answers are assembled from two ingredients: what the model already holds in parametric memory, and what it retrieves at query time from search indexes and live fetches. When a page is demoted or dropped from the index, it stops being a retrievable candidate, which means it can no longer be selected as a citation even if the assistant would otherwise have found it useful. A demotion event in classic search can quietly reduce the pool of URLs an assistant has to work with, and the effect is not visible in any organic reporting view.

The mapping is imperfect in both directions. Research into AI Overview sourcing has found that close to 30% of cited sources sit outside the first page of organic results, which means rank recovery and citation recovery are not the same job. A page can climb back into position eight and still never appear as a source. A page can be cited regularly while ranking poorly. Teams that treat rankings as a proxy for AI visibility will misread their own recovery, usually in the optimistic direction.

Platform behaviour compounds this. ChatGPT, Gemini, Perplexity, Claude and Meta AI appear to weight third-party sources differently, and each ecosystem operationalises source trust in its own way. One demotion event may materially change how one assistant describes a brand while barely touching another. That variation is the reason a single-platform check is close to worthless as diagnosis. Organic dashboards will not tell you whether recommendation share moved, because recommendation share is not a traffic metric at all; it is the frequency with which a brand appears in generated answers to the questions buyers actually ask.

Annotate 18 to 21 August 2026 in every reporting property before anything else, then resist the urge to draw a conclusion from the site-level line. Segment first, by page template, by publish-date cohort and by production method, because the site total will average a heavily affected cluster together with an unaffected one and hide both. A B2B services site with 400 near-identical location pages is the standard case: comparing impressions per cohort before and after the update usually shows the damage concentrated in the templated set while editorially produced pages hold steady.

Remember that, at the time of writing, AI Mode and AI Overview impressions are rolled into Search Console's overall Search performance data rather than reported as a separate surface, so treat any site-level impression drop as a blended figure and lean on cohort segmentation and third-party AI visibility tracking instead. Then run a structured check across the six pillars that determine whether a page is usable by both search engines and assistants: clarity, consistency, trust, visibility, freshness and technical foundations. In affected clusters the failure is rarely AI use itself. The failure is thin evidence, absent authorship and inconsistent business information repeated across hundreds of pages.

Finish the diagnosis with a live test. Build a small set of real buyer questions, phrased as a customer would phrase them, such as "which UK firms handle warehouse fire risk assessments"or "best managed IT provider in Leeds for a 60-person firm", and run them across several assistants. The purpose is to establish whether the affected pages were ever cited in the first place. If they were not, the update changed nothing about AI visibility and the remediation is a purely organic exercise. A free AI visibility scan is a reasonable starting point for establishing whether an issue exists, though a handful of prompts cannot show whether a result is repeatable across platforms.

Each affected URL has three honest outcomes, and pretending there is a fourth is how teams waste a quarter. Consolidate it into a stronger parent page. Rebuild it with proprietary detail and a named expert. Or remove it and redirect to the nearest genuinely useful page. In the 400-location-page scenario, the defensible route is usually to keep the 40 pages that carry real local evidence, such as a named office, a dated client scenario, regional pricing or a specific accreditation, consolidate the remainder into one strong parent, and redirect.

Mass noindexing the whole set is the most common overcorrection, and it is the trap nobody warns about. Programmatic pages that perform poorly in organic search are sometimes the exact URLs assistants have been retrieving, because retrieval rewards specific, well-structured answers to narrow questions rather than commercial ranking strength. Run citation attribution before deletion. Establish which pages currently appear as sources in AI answers, then decide what to prune, because a weak organic performer can still be the most cited asset a brand has in Perplexity or ChatGPT.

Something only you could have written is a concrete standard, not a sentiment. It looks like original data from your own client base, tested pricing with the date it was tested, named standards such as ISO 27001 or PAS 2035 where they genuinely apply, anonymised client scenarios with real constraints, dated screenshots, and a short methodology note explaining how a figure was produced. Technical repairs matter on both surfaces at once: confirm crawler access for AI user agents in robots.txt, correct structured data so that the entity, location and service claims match the visible page, resolve canonical conflicts created by the templated set, and rebuild internal links so authority flows into the surviving pages rather than into redirects.

Recovery inside AI answers is largely off-site work, which surprises teams used to on-page remediation. Assistants build their view of a brand from directories, review platforms, trade press, comparison and listicle pages, professional bodies and community threads, and they synthesise a description from that mixture rather than from a company's own marketing copy. If those sources are stale, contradictory or absent, a rebuilt website will not fix the answer.

The source mix is also shifting. Reddit's share of ChatGPT citations has been observed to fall from its earlier peak, which is a reminder that over-indexing on any single source type is fragile. A brand whose AI presence rests on one forum, one directory or one review platform is exposed to a change it does not control. The practical rebuilding moves are unglamorous: correct category and location data wherever the business is listed, seek inclusion in credible list and comparison pages written by people who will actually assess you, keep review profiles current, and publish research others have a reason to quote. Competitor benchmarking is the shortcut to knowing where to start, because the sources an assistant cites when recommending a rival show precisely which third parties it is borrowing authority from.

Recovery timing does not follow the fix. Spam classifier reassessment typically waits for recrawling and a subsequent update cycle, so a same-week bounce is not the realistic expectation and its absence is not evidence that the remediation failed. Plan a measurement window of weeks, and avoid stacking further large changes on top, because parallel edits make attribution impossible and leave the team unable to say which action moved anything.

Use two measurement sets, not one. On the search side, track rankings, indexed page count, impressions by cohort and the recovery curve of the surviving pages. On the AI side, track citation and recommendation frequency against a fixed prompt set, expressed as a visibility score so that movement is comparable week to week. Fifteen to one hundred real purchase questions, tested across the leading assistants and monitored daily rather than sampled once a quarter, is enough to make a drop visible within days of an algorithm event. The AI visibility plans and pricing page sets out how that monitoring cadence is structured in practice.

Before-and-after evidence is what turns monitoring into proof. Retest the same prompts after each discrete fix, in the same phrasing, and record the answer text as well as the sources. Correcting a single brand miscategorisation in a widely referenced directory, for example, can change how assistants describe a company, and retesting the prompt set is the only way to demonstrate that the change was responsible. Where a deeper diagnostic is needed across multiple platforms and competitors, an advanced AI visibility audit covers source analysis and competitor benchmarking alongside the technical review.

A pre-publish gate is more effective than any post-hoc audit, and it takes four questions. Who is the named author and what makes them qualified to write this. What is the source of each factual claim on the page. What first-hand element does this page contain that no competitor has. What does this add that the current top results do not. A page that fails any of the four does not publish. That gate is cheap, and it removes the ambiguity that leads to hundreds of near-identical URLs.

Freshness is the pillar most often misread, usually as permission to publish more. Freshness means the pages that matter carry current, verifiable detail: updated prices with the date they were checked, current standards, current team, current case evidence. High-velocity thin publishing is what the update punished, and it also dilutes the clarity and consistency signals AI systems use to form a view of a brand, because forty pages saying slightly different things about the same service produce an uncertain answer.

For governance, the four functions of the NIST AI Risk Management Framework, Govern, Map, Measure and Manage, transfer cleanly to content production rather than models alone: define who owns publishing standards, map where AI is used in the workflow and what evidence each stage adds, measure output against the gate and against both search and AI visibility, and manage the corrections on a fixed cycle. A quarterly content audit rhythm catches decay, duplication and drifting business information before an update finds them. Teams weighing up how much of this to run internally can request further information on the approach before committing to a programme.

What did the August spam update target in AI content?

It targeted scaled content abuse as defined in Google's spam policies, meaning large volumes of pages produced primarily to rank rather than to help a reader, alongside site reputation abuse. The recurring characteristics in affected sites were templated page families, sharp increases in publishing velocity, near-identical introductions, missing authorship and no first-hand evidence. AI involvement in drafting was not itself the trigger.

Does Google penalise all AI-generated content, or only mass-produced pages?

Google's published policies address how content is produced at scale and to what purpose, not whether a language model assisted the drafting. Publishers running AI-assisted workflows with genuine subject-matter review, named authors and original data reported little or no movement during the August rollout. The practical test for each page is whether it contains something that could only have come from your business.

Can a Google spam update change whether ChatGPT or Perplexity recommends my brand?

It can, because grounded answers depend on retrieval from search indexes and live fetches, so pages that are demoted or dropped stop being candidates for citation. The effect is uneven, since assistants appear to weight third-party sources differently, and one platform may change noticeably while another does not. Testing the same buyer prompts across several assistants is the only reliable way to see where the impact landed.

How long does recovery from the August spam update usually take?

Reassessment normally waits for recrawling and a later update cycle rather than responding to the fix directly, so a measurement window of weeks is the realistic planning assumption. Avoid making further large structural changes during that window, because overlapping edits make it impossible to attribute any subsequent movement. Recovery in AI answers may follow a different timeline again, since it also depends on third-party sources being updated.

Should I delete pages hit by the update or rewrite them?

Check citation attribution before deciding, because a page with weak organic performance can still be a frequently cited source inside AI assistants, and removing it is a loss that no ranking report will show. Where a page has no distinguishing evidence and no citation record, consolidation into a stronger parent with a redirect is usually the better outcome. Where the underlying topic matters commercially, rebuilding with proprietary detail and a named expert is preferable to removal.