Why Google Killed CNET's AI Article Strategy (And What We're Doing Instead)

Between November 2023 and January 2024, three high-traffic publishers (CNET, Sports Illustrated, and Bankrate) were each publicly exposed for publishing AI-generated articles with fake or undisclosed author bylines. CNET's AI articles contained factual errors that ran uncorrected for weeks. Sports Illustrated's AI-generated authors had headshots traceable to a stock-photo AI service. NPR reported on the Sports Illustrated scandal as the story broke in November 2023. Here is what we know about why, and how the gotcontext.ai /news section is built to avoid the same outcome.

The penalty wave: CNET, Sports Illustrated, Bankrate¶

CNET

CNET began publishing AI-generated financial explainer articles in late 2022 under the byline "CNET Money Staff." In January 2023, Futurism found that at least 41 articles contained factual errors (incorrect interest rate calculations, wrong product information, outdated figures. CNET paused the program, issued corrections, and later removed many of the articles. Organic search traffic to CNET's money content declined substantially, a drop that coincided with Google's March 2024 Helpful Content Update.

The mechanism was not a specific "AI content" penalty. The March 2024 Helpful Content Update and the Scaled Content Abuse policy targeted pages that demonstrated no original expertise, contained factual errors, and showed evidence of high-volume templated generation. CNET's articles fit all three criteria.

Sports Illustrated

In November 2023, Futurism revealed that Sports Illustrated had published AI-generated product reviews under fabricated author names ("Drew Ortiz," "Sora Tanaka") with headshots purchased from an AI-image stock service. The author profiles had no digital footprint. SI deleted the articles within days and terminated its vendor contract. The lesson from Sports Illustrated is not that AI content is penalized. The lesson is that fake bylines are a zero-tolerance violation.

Bankrate

Bankrate's case is more instructive because the outcome was different. A Rankability case study found that Bankrate uses AI drafts combined with subject matter expert review and explicit "AI-assisted" disclosure labels, and still ranks in the top 3 for highly competitive financial terms. Articles that carried transparent AI labels and real expert review continued to rank. Those that did not suffered. Bankrate is the clearest proof in the public record that AI-assisted content is not inherently penalized. Process transparency and human expertise review are what differentiate acceptable from flagged.

What Google actually penalized¶

Reading the pattern across all three cases, the trigger was not "AI was used in the writing." Google's March 2024 core update and spam policies documentation states that the Scaled Content Abuse policy targets content generated "at scale" that "doesn't demonstrate enough value or originality" regardless of whether a human or an AI produced it. Three specific patterns recurred across all three properties:

Fake bylines and AI-generated author photos. This is the most direct E-E-A-T violation. An author who does not exist, or whose photo resolves to an AI image generator, fails the trust check entirely. SI's "Drew Ortiz" returned zero results in any search engine.
Factual errors a real editor would have caught. CNET's articles contained arithmetic errors. A subject matter expert reading the draft for five minutes would flag them. The errors became evidence that no review process existed.
Scaled content with no original expertise signal. The Scaled Content Abuse policy fires on high volume combined with thin, templated output. CNET published at a rate implying a large team of financial writers. The discrepancy between publication rate and visible editorial infrastructure is itself a signal Google's systems detect.

There is a fourth risk specific to our situation. Google's Site Reputation Abuse policy (May 2024) penalizes sub-folder content that does not match the host domain's established expertise. gotcontext.ai ranks today for technical queries: compression ratios, MCP tooling, context engineering. A flag on /news/* can cascade to /docs and /pricing, the surfaces the product depends on. That cascade risk drove our Variant B choice before we even factored in content quality.

Variant A vs B vs C¶

The /news architecture document at docs/architecture/AI-NEWS-AGGREGATOR.md §2 names three implementation choices:

Variant A: full auto-publish. Poll sources hourly, LLM-draft articles, publish with no human gate. This matches the CNET pattern precisely. The architecture doc rates SEO risk as HIGH and recommends against it on the main domain even with subdomain isolation, because the factual-error and thin-content risks are structural.
Variant B: LLM-draft with human approval. Same ingestion and drafting pipeline, but articles land in a pending_review queue. A human editor approves, edits, or rejects before anything publishes. This is what The Decoder and most AI-news sites with sustained traffic actually do. SEO risk: LOW, because every article carries a real byline and a reviewed-at timestamp.
Variant C: hybrid tiered. Auto-publish for structured, low-stakes content (link roundups, daily digests). Human gate for opinion pieces and analysis. The documented evolution path after Variant B shows engagement signal over 4-6 weeks, per AI-NEWS-AGGREGATOR.md §13 Phase 4.

We chose Variant B for Phase 1. The reasoning from AI-NEWS-AGGREGATOR.md §11.2 is direct: for Variant B, the canonical link on re-written articles points at the source URL, voluntarily conceding ranking to the primary publisher while keeping the topical traffic benefit. Variant A with subdomain isolation is buildable, but the asymmetric downside (a whole-domain deindex from a cascade penalty) is not worth it for a small team where the main domain's search authority is the product moat.

What we do differently¶

Six concrete choices in the /news pipeline address the documented failure modes. These are architectural decisions in code and configuration, not editorial policies that drift.

Real bylines, no fabricated authors. Every published article carries a real contributor name. No stock photos, no AI-generated headshots, no author who does not exist. The Sports Illustrated failure was entirely about this. The policy is codified in project rules and enforced at the review queue.
Canonical URL stays at the source. Per AI-NEWS-AGGREGATOR.md §11.2, on per-article pages that primarily re-write a single source, we set <link rel="canonical" href="{source_url}">. This voluntarily concedes the ranking on that page to the original publisher.
Every draft goes through a human gate before publish. The drafting pipeline sends each LLM-generated article to a review queue with status pending_review. Nothing reaches a public URL until a human has read it. This is the factual-error check that CNET's pipeline skipped.
Explicit AI disclosure on every article. Every published piece carries a visible label: "Drafted by AI, reviewed by [editor name] on [date]." Transparent disclosure is neutral-to-positive for E-E-A-T. Concealing the process is what got CNET and SI penalized, not the process itself.
Prominent outbound attribution. The original Reddit thread, arXiv paper, or HN discussion is linked prominently at the top and bottom of every article. We are adding the engineering-team perspective that the primary source does not have.
Rate-limited ingest at 1 article per cycle. High volume combined with thin content is exactly the Scaled Content Abuse pattern. A slow cadence with high editorial density is the opposite signal.

What we have not tested¶

Honest disclosure

The gotcontext.ai /news section has zero published articles as of 2026-05-20. The architecture is built. The pipeline is wired. The SEO outcome is not yet measured.
We do not have CNET-scale traffic to validate our own thesis. We cannot yet show a Google Search Console graph where Variant B holds rankings that Variant A would have lost. That data is 90+ days away at minimum.
The Variant B architecture is designed to avoid the documented failure modes from CNET, SI, and Bankrate. We are claiming that the design avoids them. We are not claiming that we have proven our approach works in production.
Google's March 2026 E-E-A-T update eliminated 24.1% of Top-10 pages in the most volatile update in recent search history. Its full impact on AI-news-specific sites is not yet fully documented. We will check Google Search Console manually every week for the first 90 days.

Try the /news section¶

The /news section is live at gotcontext.ai/news. Today it contains 38 hand-curated items from the static seed that shipped with PR #114 in May 2026. The dynamic pipeline (ingestion, LLM drafting, review queue, human-approved articles) is being wired in Phase 1. The first batch of drafted articles will appear in the queue before they appear on the site.

If you visit /news today, you will see a clean feed of curated links and no long-form articles. That is the honest production state. We built the SEO posture into the architecture before we built the content volume, the opposite of what CNET did.

What's next¶

The next post in this series covers the engineering side: how the ingestion pipeline works, what the LLM drafting prompt looks like, and how the review queue wires into the existing gotcontext API. A later piece covers self-hosting the pipeline for teams that want to run it against their own source list.