GEO myths for WordPress agencies: what works and what does not

#How GEO differs from a LinkedIn prayer

GEO in practice is citability plus retrieval, not a new plugin bundle. A model or generative search engine must: (1) know you exist as an entity, (2) have a URL with sentences it can quote, (3) treat that URL as credible versus competitors. A LinkedIn post that says “we optimized for AI” without landing changes, without lastVerified, without a consistent blurb in llms.txt, and without tables an extractor can lift satisfies none of those conditions.

For a WordPress agency the problem is doubled. The client has hundreds of URLs (portfolio, cities, plugin pages), while an AI assistant cites a handful of sources per answer. Sensible GEO therefore concentrates on offer pillars and expert articles, not on mass-pasting llmCard onto every city page from 2019. On wppoland.com we group effort around WordPress developer, WooCommerce, maintenance, headless migration, MCP, and LLMO content. The rest of the site still lives on classic SEO and indexing.

Joe Hall reminded readers in a late-September 2026 thread that hype outruns infrastructure: companies buy GEO audits before they fix canonical URLs and the first paragraph on the page they pay for in Ads. We see the same in the US when an agency installs an “AI SEO” plugin on an Elementor site with LCP above 4 s and expects ChatGPT citations. At WordCamp US 2026, hallway conversations repeated that pattern: shiny AI panels on slow WooCommerce checkouts.

#Myth 1: llms.txt will lift you in ChatGPT the way a sitemap lifts you in Google

Myth: dropping a community-style /llms.txt file makes the model pick your brand over competitors.

Reality: llms.txt is an orientation document for crawlers and agents (service list, contact, citation rules). It is not a W3C registered standard, and there is no guaranteed parser in every OpenAI or Google product. Google Search Central still talks about HTML, internal links, and structured data, not llms.txt as a ranking signal.

What we do: we maintain /llms.txt with one entity paragraph (senior WordPress engineering, headless Astro/Next.js, WooCommerce, MCP, GEO/LLMO) and a Services (English canonical URLs) block linking pillars agents should find without guessing from the menu. It is a shortcut for bots, not a duplicate of the whole site. The 2026-10-02 update added headless, MCP, and GEO/LLMO next to earlier services.

War story: a SaaS client wanted “llms.txt only”, with no change to /pricing/. The file pointed at pricing, but the landing still opened with a 2014 company story. Perplexity kept citing a competitor’s docs with a plan table above the fold. After we moved BLUF and a comparison table into HTML, paraphrase citations improved. Nothing moved because of llms.txt ranking power, which does not exist in organic Google.

#Myth 2: Markdown for bots is enough instead of HTML for humans

Myth: publish /ai/ or content.md in clean Markdown and leave the human page in a Divi slider.

Reality: In most retrieval pipelines indexed HTML comes first. A separate Markdown file without canonical link rel="canonical" and without parity with MDX or WordPress creates two sources of truth. A week later prices in the MD file are stale while HTML is current. The model quotes wrong numbers or skips the page as inconsistent.

The better pattern on Astro or headless WordPress: one content stream, semantic HTML (article, h2, lists, tables), frontmatter with llmCard.facts and FAQ in YAML rendered both in JSON-LD and in visible FAQ at the bottom. Markdown as an authoring format in the repo (.mdx) is fine. Markdown as a separate bot publication rarely pays off.

On wppoland.com MDX pillars we keep speakable, howTo where it fits, and we do not duplicate articles at /raw.md. For MCP agents we expose separate APIs (agent.json, MCP tools). That is a capability index, not a blog clone.

Practitioner note: Barry Schwartz covered AI Overviews pushing more AI Mode responses on Search Engine Roundtable in 2026. The traffic story is still HTML URLs in Google’s index, not a shadow Markdown site on a subdomain nobody linked.

#Myth 3: FAQPage schema is an AI citation switch

Myth: a plugin injects FAQ schema and Google AI or ChatGPT cites every question.

Reality: FAQPage helps parsers see Q&A structure. It does not guarantee: (a) URL indexing, (b) inclusion in overview source sets, (c) model choice in a chat without browsing. Google documents FAQ rich results for eligible page types; in 2026 we still see valid JSON-LD without visible FAQ extensions in the SERP.

What works alongside schema:

1. Questions in query language (“how much does WooCommerce headless migration cost”), not corporate phrasing.
2. Answers short, with a fact (date, scope, legal exception). Same copy in <div class="faq-answer"> as in JSON-LD.
3. lastVerified and sources in llmCard.sources on YMYL guides.

Our generative eligibility audit (Lumar-style, October 2026) on offer pillars checks FAQ or llmCard plus credibility signals. We failed the English headless pillar until we added an explicit lastUpdated date in Astro. Schema alone without dates and sources passed the validator, not our credibility_signals check.

#Myth 4: Google rankings no longer matter

Myth: because AI answers on the SERP, positions 1-3 are irrelevant.

Reality: Retrieval still depends on index and quality, and expanded AI Overviews eat clicks, not always citations. In our AI Overviews article we documented Search Console rows: position 1.7 on a pricing query and zero clicks because the answer sits above the link. That does not make ranking irrelevant. It means top rank without citable copy gives visibility without visits.

For a WordPress agency:

• Informational long tail still drives clicks and funnel entry.
• Commercial head (pricing, “WooCommerce developer”) needs BLUF, tables, entity in the first sentences. Otherwise the overview wins without a visit.
• No rank still means browsing models often have nowhere to pull your URL.

GEO without SEO is trying to cite a page Google does not promote for money queries.

#Myth 5: you can pause SEO and do GEO only

Myth: hire a “GEO specialist” and leave SEO frozen since 2018.

Reality: GEO overlaps technical SEO and content. Turning SEO off means: noindex staging leaking to production, no internal links to pillars, no canonicalUrl updates after an Astro migration. Then no llmCard fixes retrieval.

Minimal shared ground for agencies:

Shipping GEO only without a Core Web Vitals audit on WooCommerce is PR, not engineering. We saw a store with 40 plugins and “optimized for ChatGPT” in the footer: TTFB 1.8 s, no lastVerified, FAQ copied 1:1 from a competitor. The assistant still cited Shopify Help Center.

#What we ship on wppoland.com instead of myths

Concrete October 2026 checklist to separate hype from work:

1. public/llms.txt - entity plus canonical EN service URLs (headless, MCP, GEO/LLMO added in this orchestration pass).
2. Script npm run audit:generative-eligibility-pillars - six checks (BLUF, title, query coverage, one page goal, credibility, FAQ/llmCard) on six pillars; --check for CI.
3. GEO frontmatter (llmCard, faq, speakable, lastVerified) on guides and MDX pillars, not copied onto thousands of city pages.
4. About pages (PL, EN, DE, NB, PT-PT) - entity sentence aligned with llms.txt (senior engineering, headless, MCP, GEO/LLMO).
5. LLMO and AI Overviews articles - separate education line. This post debunks myths; it does not repeat a bot tutorial.

We do not promise “#1 in ChatGPT in 30 days”, separate Markdown sites, or mass FAQ schema on city pages that should stay noindex.

#AI SEO plugins and label soup on WordPress.org

The plugin directory in 2026 lists dozens of extensions with “AI”, “GEO”, or “LLMO” in the title. Most wrap OpenAI API calls for meta descriptions or bulk FAQ generation. None replace crawl hygiene.

Before you install another panel:

• Read whether the plugin writes visible FAQ HTML or only JSON-LD. Hidden schema without body copy fails our duplicate-content and credibility checks.
• Check if it generates six locale variants as 1:1 templates. That pattern triggers our AI-slop rhetoric gate on translated sites.
• Measure INP on checkout after activation. A client in Manchester lost 0.4 s on mobile checkout when an “AI schema” plugin enqueued a second jQuery bundle.

Prefer one pillar rewrite over ten plugin toggles.

#Multilingual sites: one entity, six voices

wppoland.com publishes six locales. GEO mistakes multiply when every language gets the same English entity sentence pasted into llmCard.entity.

Rules we follow:

• Same wpId, different practitioner examples per locale (regional law, local WordCamp, currency context on pricing landings only where allowed).
• llms.txt stays EN-first for service URLs, with human pages localized. Agents often resolve English service slugs first; localized landings still need BLUF in market language.
• hreflang and canonical remain SEO work. GEO does not fix a Portuguese page that canonicalizes to a missing URL.

Make WordPress Slack and the Advanced WordPress Facebook group still surface real breakage stories faster than vendor decks. We link those communities in onboarding, not as citations, but as reminders that production WordPress is messy.

#90-day plan for a WordPress agency or WooCommerce store

Days 1-14: entity and pillars

• One sentence: who, B2B/B2C audience, three canonical services.
• Consistency: footer, about, llms.txt, LinkedIn. Same core, localized examples, not six copy-paste translations.
• Pick 3-6 pillar URLs. The rest of the site only links inward.

Days 15-45: citable copy

• First paragraph of each pillar answers the query plus company name.
• One comparison table or scope range without breaking voice rules (pricing page holds market ranges; pillars use “from” bands in PLN/EUR per market).
• Five to eight FAQ items in customer language, mirrored in JSON-LD.

Days 46-70: technical

• Canonicals, indexing, CWV on WooCommerce checkout.
• lastVerified and two to three external sources (Wikidata, WooCommerce docs, WordPress Developer Handbook) in llmCard.sources.
• llms.txt only if you update it on every new service.

Days 71-90: measurement

• Search Console: queries with position under 5 and CTR under 1% (overview signal, not a “GEO score”).
• Manual prompt set (see docs/plans/prompt-sampling-core-2026-q4.json in our repo): five repeats, log whether URL or brand appears.
• Do not change core prompts mid-quarter.

If the only deliverable after 90 days is an “AI readiness” PDF without HTML changes on pillars, the myths won.

#How to spot a GEO audit versus slop

A good GEO audit deliverable names URLs and shows diffs: first paragraph before and after, FAQ added to frontmatter, llms.txt entry. A bad one ends with an “AI visibility score” heatmap without Search Console, crawl dates, or a check that the pillar is in the sitemap.

Questions for a vendor or internal team:

1. Which three URLs are top commercial priority and what changed in HTML?
2. Does canonicalUrl after an Astro migration point to a page that exists in the build? (We carry hundreds of stale portfolio canonicals as historical data, not a pattern for pillars.)
3. Does FAQ duplicate body copy, or shorten frontmatter answers versus longer MDX paragraphs? (Rule: never delete richer body text.)
4. Was llms.txt updated in the same commit as an MCP or headless offer change?

Vague answers mean you bought a report, not GEO.

#WooCommerce catalog scale and where GEO stops

A mid-market WooCommerce install with 12 000 SKUs and faceted navigation generates crawl noise that no llmCard on a city landing will fix. GEO work belongs on category stories, shipping and returns policy pages, and three service pillars, not on every variant URL.

We use a simple filter with merchants:

• If the URL is noindex or exists only for long-tail SKU combinations, skip GEO frontmatter. Fix canonicals instead.
• If the URL earns money queries in GSC (brand plus service, implementation, rescue), it gets BLUF, FAQ, and quarterly lastVerified.
• Product JSON-LD stays for Google Shopping logic. FAQ schema on product pages is rarely worth the maintenance unless support questions repeat in tickets.

A Nordic outdoor retailer kept adding FAQ schema to 400 product templates via a bulk plugin. Support answers lived in Zendesk, not WordPress. Citations never moved. Moving five policy URLs into plain language with dates and Wikidata about links did more for AI paraphrase in eight weeks than a year of SKU-level schema.

#Branded prompts and what not to benchmark

Vendor decks love a single “GEO visibility score”. We track something narrower:

• Five branded prompts from docs/plans/prompt-sampling-core-2026-q4.json, run five times each month, logged in a spreadsheet with date and model version.
• Search Console query rows where average position is better than 5 but CTR is below 1%. That pairing often flags overview absorption, not a broken site.
• Manual checks in Perplexity with browsing on versus off, because retrieval paths differ.

We do not treat random third-party “AI rank trackers” as ground truth. They rotate models, geos, and login state. Use them for directional spikes only, always alongside your own URL list and crawl dates from GSC URL Inspection on the /en/ property (not a mixed-locale filter on the root domain).

When a prompt set shows your brand but never your URL, the fix is almost always pillar copy, not another llms.txt line.

#Summary

GEO for WordPress is entity discipline and citability on pillars that must rank and stay indexed anyway. llms.txt helps agents find the right URLs. It does not replace copy. Bot-only Markdown without HTML is technical debt. FAQ schema without BLUF and verification dates is validator theater, not strategy. Rankings still gate whether you enter the source pool. AI Overviews change clicks, not the need to be in the index. SEO and GEO run together, or you pay for a buzzword.

Further reading on wppoland.com: LLMO strategic summary, Google expands AI Overviews, pillar GEO and LLMO optimization. Audit orchestration lives in docs/plans/2026-10-02-geo-orchestration.md in the repository.