When we started building BD Auto Blog, the obvious choice was OpenAI’s GPT. Bigger ecosystem, more tutorials, more existing WordPress integrations. We picked Anthropic’s Claude API instead. This post explains why, what it means for the content the plugin produces, and how we built guardrails to keep you out of Google’s scaled-content trap.
If you just want the plugin, BD Auto Blog is here. If you want to understand what’s actually happening before you trust it with your site, read on.
Why Claude over GPT
A few specific reasons, none of them about which model is “smarter.”
1. Editorial voice consistency. In our testing, Claude held a defined editorial voice across long articles better than GPT-4 of the same era. Drift inside a single 1,500-word piece — switching from conversational to formal mid-section, or losing the “speak directly to the reader” instruction — was less common. For a plugin that’s writing under your brand name, voice consistency is the single most important thing.
2. Refusal behavior on weak claims. Claude pushes back when you ask it to make claims it can’t substantiate, more often than GPT does. For a content plugin that publishes to your live site, a model that’s more conservative about fabrication is the safer ground floor.
3. Longer effective context for outline-then-draft workflows. Our generation pipeline is two-pass: generate a structured outline first, then draft the body against that outline. Claude’s context window comfortably holds the outline, the brand voice instructions, and the draft in one prompt without truncation tricks.
4. Pricing predictability for our cost model. API pricing for Claude Sonnet 4.5 (our default model, claude-sonnet-4-5) sits in a range that makes our license tiers viable. GPT-4 class models cost more for comparable output length. We pass the API cost transparently — you bring your own Anthropic API key, and we store it encrypted (AES-256-CBC) in your WordPress install.
None of these are knockout arguments. GPT works fine. We picked Claude because the specific tradeoffs aligned with what we wanted to ship. If Anthropic raises prices or OpenAI ships a model that’s clearly better at our use case, we’ll re-evaluate.
How the prompt structure works (generic)
We won’t share the exact system prompt we use — that’s a real piece of IP and shipping it publicly would let it be reverse-engineered or undercut by every competitor in a week. But here’s the structure, generically:
- Role and constraints — the model is told it’s writing for a specific publication, with a specific voice, for a specific reader.
- Topic and outline — generated in a separate API call, passed in as structured input.
- Voice rules — explicit “do not use” list (the AI tells: delve, leverage, unleash, harness, robust, comprehensive, seamless, cutting-edge, etc.) and a “prefer” list (specific numbers, active voice, short sentences).
- Output schema — H2 structure, paragraph length targets, internal-link slot markers.
- Quality gates — instructions to flag (not suppress) any claim it can’t substantiate, so our warning system catches them post-generation.
The output goes through our warning system before it ever hits your WP database. That’s the part that matters more than the prompt.
The warning system
Every AI generation tool produces hallucinations. Anyone telling you otherwise is selling you something. The question is what your pipeline does about them.
BD Auto Blog flags posts at generation time when it detects:
- Fabricated statistics — patterns like “X% of users” or “studies show” without an attached source link
- Specific company / product claims — mentions of named third parties making specific claims, which is where defamation risk lives
- Time-sensitive claims — “as of 2026” type statements that age quickly
- Image-text mismatch — when the featured image’s Pexels metadata doesn’t match the article’s subject closely enough
- Length anomalies — short paragraphs, missing H2s, or generated content under the target word count
- Repeated phrasing — when the model fell into a refrain or repeated a sentence structure too often
Flagged posts land in your queue with a warning indicator (amber dashicons-warning in the admin list) and a JSON-encoded warning array attached as post meta (_bdab_warnings). The post does not publish until you clear or accept each warning.
This is the thing that separates a publishable AI-blog plugin from a liability. Anyone can pipe a model’s output to wp_insert_post(). The work is in the validation layer.
Pexels images vs synthetic AI images
We use Pexels for featured and inline content images. Not Midjourney. Not DALL-E. Not Stable Diffusion.
Three reasons:
1. They’re real photos. Real photos of real things signal authenticity to readers in a way that the slightly-off-aesthetic of mid-2026 AI image generators doesn’t. Hands look right. Text in the background reads as actual text. Pexels images don’t have the uncanny-valley signal that’s increasingly trained into reader instincts.
2. Attribution is baked in. Every image we pull stores Pexels-side credit data as post meta (_bdab_pexels_credit, _bdab_pexels_url, _bdab_pexels_id). You can render the credit at the bottom of the post or in the image caption. Synthetic images skip this entirely, which is also why some are starting to fall foul of provenance metadata requirements.
3. Licensing clarity. Pexels images are free to use commercially with optional attribution. We pull only from Pexels’ verified contributors. With synthetic images, the copyright situation is still unsettled — particularly for training-data lawsuits making their way through US and EU courts. We didn’t want to put you in the middle of that.
Content images on the Professional tier insert before H2 sections, sourced from Pexels via topic-relevant search. The default is one image per 800 words of body content, configurable.
License-gated features
Three tiers:
- Starter ($49/yr) — Topic generation, scheduled posting, basic article writing. Limited posts per day (capped at the lower of your settings and the tier ceiling).
- Professional ($99/yr) — Adds Pexels featured + content images, social snippet generation, CTA blocks.
- Agency ($199/yr) — Adds white-label option (your branding in the admin UI, not ours), unlimited posts/day, weekly digest emails to clients.
The full breakdown lives on the pricing page. The All-in-One bundle includes BD Auto Blog Professional alongside BD Security Firewall, backup, malware scanner, and the rest of our ten plugins.
Feature gating is enforced server-side at multiple points — scheduler, article writer, helpers, admin UI. Bypassing the license key doesn’t enable gated features (we built that pipeline deliberately).
The SEO risk you need to understand
In March 2024, Google updated its spam policies to specifically target “scaled content abuse” — the practice of generating large volumes of content “primarily to manipulate search rankings rather than to help users.” This was widely (and slightly incorrectly) reported as “Google penalizes AI content.”
The actual policy is more precise. Google’s update says, paraphrased: scaling content production isn’t inherently against guidelines. Scaling low-quality content production for the purpose of manipulating search is. The distinction matters.
What gets you penalized:
- Topic strings unrelated to your site’s expertise — generating posts about cryptocurrency on a gardening blog
- Volume disconnected from quality control — publishing 50 AI posts a week with no human review
- Templated structure that’s obvious to readers — same intro pattern, same conclusion pattern, same H2 sequencing across every post
- Fabricated facts that get flagged on review — Google does sample-check
- Thin content with no actual answer to the search query — the model padded to length without substance
What’s safe — and what BD Auto Blog is designed to support:
- Topics within your site’s actual subject area — the topic generator constrains by your configured site focus
- Reasonable cadence — even on Agency tier, we recommend not exceeding 1–2 posts per day on a single site
- Human review on every post before publish — the default workflow stages posts as drafts; the warning system requires sign-off
- Voice consistency that reads as your brand — see the prompt structure section above
- Fact-checked claims — warnings flag the dangerous ones
Used as a draft-generator that you edit and publish, BD Auto Blog reduces the time-cost of WordPress blogging meaningfully. Used as an unattended firehose, it can hurt you. That’s true of every AI content tool. We’ve built ours to nudge you toward the first use case.
What we will not claim
We won’t tell you AI-generated posts will rank. Ranking depends on your site’s overall authority, the competitive density of the query, and whether the post genuinely answers the search intent. The tool produces drafts; you do the rest of the SEO work.
We won’t tell you our warning system catches every hallucination. It catches patterns we’ve seen. Novel hallucinations slip through. Human review is required, not optional.
We won’t tell you Claude is universally better than GPT. It’s the right choice for our specific pipeline. Your mileage with a different model could differ.
Try it
If you want to test BD Auto Blog with your own Anthropic API key on one site, start at the product page. Starter tier is $49/yr. 30-day refund if it doesn’t fit your workflow.
If you want the full BD plugin catalog including BD Auto Blog Professional, the All-in-One bundle is on the pricing page.
FAQ
Will Google penalize AI-written posts?
Not for being AI-written. Google’s March 2024 spam policy targets scaled production of low-quality content meant primarily to manipulate rankings. AI-assisted posts that are within your site’s topic area, reviewed by a human, and substantively answer the search query are not violations of that policy. AI posts published unattended at high volume with no quality control are at risk.
Why Claude over GPT-4?
For our specific pipeline: better editorial voice consistency across long articles, more conservative behavior on unsubstantiated claims, predictable pricing for our license tiers. None of this means Claude is universally better. GPT works fine for many content workflows; it wasn’t the right fit for ours.
Can I edit posts before they publish?
Yes, and we recommend it. The default workflow stages every generated post as a WordPress draft. The warning system blocks publish until flagged items are reviewed. You can also turn off auto-publish entirely and use BD Auto Blog as a pure draft-generator.
What happens if Claude generates inaccurate facts?
Our warning system flags patterns associated with hallucination (unsourced statistics, named third-party claims, time-sensitive assertions) before publish. The post stays as a draft until you clear or accept each warning. The system isn’t perfect — novel hallucinations can slip through. Treat every generated draft as needing fact-check, not as final copy.
Does my API key stay on my site?
Yes. Your Anthropic API key is stored encrypted (AES-256-CBC) in your WordPress options table. We don’t have access to it. Generation requests go directly from your WordPress install to Anthropic’s API — they don’t route through any BD-hosted service.
