Running SEO for one WordPress site with AI help is a solved problem. Running it for six distinct brands, in English and Arabic, without the output collapsing into interchangeable sludge — that is an operations problem, not a prompting problem. This is how I built Claude SEO Squad: 12 agents covering research, content and optimisation across 6 WordPress properties, and what actually broke along the way.
The problem: six brands, one team, two languages
Every WordPress brand needs the same pipeline: keyword research, competitor monitoring, content briefs, drafting, editorial review, publishing, and on-page maintenance. Do that manually for one site and it fills a content person's week. Do it for six and the work does not scale linearly — it scales worse, because context-switching between brands is its own tax. Each property has a different audience, a different topical focus, and a different tolerance for tone. Then double most of the content burden again for Arabic, which the GCC market demands and which most SEO tooling treats as an afterthought.
The single-site AI SEO tools — Surfer, Clearscope, the Jasper tier — assume one brand, one language, one operator in the loop. None of them model the coordination problem: which brand gets which keyword, how briefs stay distinct, and who catches the Arabic draft that reads like machine translation.
The 12-agent breakdown
The squad decomposes the pipeline into single-responsibility agents rather than one mega-prompt. The research layer: a keyword-hunting agent that mines search data for opportunities and maps each one to exactly one brand, and a competitor-tracking agent that watches what ranks in each niche and flags content gaps. The production layer: a brief-generation agent that turns an approved opportunity into a structured outline with intent, angle and internal-link targets; drafting agents that write against the brief; and a curation agent that decides what is worth producing at all. The quality layer: an editorial review agent that gates every draft against brand voice, factual claims and SEO fundamentals before anything approaches WordPress.
The important split is shared logic versus brand-specific configuration. Research, brief structure, QA criteria and publishing mechanics are shared infrastructure — one implementation, six deployments. Voice, topical boundaries, banned phrases and audience assumptions live in per-brand configuration files that every agent loads before it touches that property. You do not need 72 agents for 6 brands; you need 12 well-factored ones with brand context injected at runtime.
Keeping six brands from sounding like one
Sameness is the default failure mode of scaled AI content. Left alone, a model regresses to its house style and every brand ends up publishing the same article with different logos. Three things prevent that here. First, each brand's configuration is written from real published material — actual sentences the brand has used, not adjectives like "friendly but professional". Second, the keyword agent enforces topical separation: when two brands could plausibly target the same UAE query, one gets it and the others are explicitly blocked from it, which also kills cross-brand cannibalisation before it starts. Third, the review agent scores drafts against the specific brand profile, not against generic quality, and rejects anything that could have been published on a sibling site unchanged.
ONE TACTIC A WEEK
The bilingual pipeline: Arabic is not a translation step
The early version translated finished English posts into Arabic. It failed review constantly. Translated content carries machine-translation tells — English sentence rhythm, idioms rendered literally, examples that make no sense to a Gulf reader — and it inherits English-first keyword targeting when Arabic search behaviour for the same intent is often different. The current pipeline generates Arabic natively from the brief, with Arabic keyword research done separately, so the draft is written for the Arabic reader rather than converted for them.
The honest limit: automated QA catches structural problems — dialect inconsistency, awkward calques, missing diacritics where they matter — but a native-Arabic human reviewer still signs off before publish. I will not claim full automation there, because cultural fit is exactly the kind of judgment that fails silently. The agents made the reviewer's job smaller, not obsolete.
WordPress integration
Publishing goes through the WordPress REST API — drafts are pushed with title, slug, category, meta description and internal links already in place, but they land as drafts, not published posts. The review-then-publish step is deliberate: an agent with unattended publish rights across six brands is a liability, not a feature. Internal linking is handled per brand from a maintained map of each property's URL structure, so links are chosen from a known-good set instead of hallucinated. On-page elements — titles, headings, schema-relevant structure — are validated by the QA agent against the brief before push.
Results
The system runs in production today: 12 agents across 6 brands, producing bilingual content on a continuous cycle. I am deliberately not publishing traffic and ranking numbers for the squad's own properties yet — those need more months of data before they mean anything, and I would rather report them late than dress them up early. The closest audited SEO outcome in my client work is Rainbow Printing: +50% organic traffic and +27% conversion rate. That was a separate engagement, not the squad's output — but it is the standard the squad is built to hit.
Lessons and tradeoffs
Three lessons. Quality control fails at the review layer first: when volume rises, the temptation is to loosen the QA gate to keep throughput, and that is precisely when you start shipping spam. The gate stays fail-closed — a rejected draft dies, it does not get waved through. Second, LLM cost at this scale is real but it concentrates in drafting and review, so cheaper models handle research and classification while the stronger model is reserved for writing and QA. Third, the human editor is non-negotiable at two points: Arabic sign-off and anything making a factual or price claim. Everything else the machine handles better than a tired human on their sixth brand of the day.