How we tracked wppoland.com's own AI citations for 90 days

Most GEO content stops at tactics. Procurement teams need something else: a reproducible baseline, declared limitations, and a cadence they can map to vendor reviews. We already published a Q2 2026 measurement snapshot that showed how much the instrument changes the story. This series narrows the question: if we hold prompts constant for ninety days, does off-page authority work move transactional citations, or does identity stay strong while money queries stay empty?

That is a programme question, not a blog gimmick. We answer it in public because we sell GEO and LLMO optimization and we refuse to recommend spend without a traceable counter.

#Query families we track

We split prompts into families on purpose. Mixing them produces a single “visibility score” that flatters identity wins and hides transactional losses. Procurement should reject that composite.

#Identity and positioning prompts

These are narrow, branded questions a buyer asks when they already know Poland is in scope. Our anchor prompt, the one in the Geoboard baseline, is effectively: Polish agency for foreign WordPress clients. Variants include nearshore framing, EU timezone delivery, and English-first communication for US or UK stakeholders.

This family measures whether models resolve who you are, not whether they would hire you for a six-figure WooCommerce rebuild. Winning here is necessary and insufficient.

#Informational and practitioner prompts

These mirror how engineers and leads research before RFP language appears: headless WordPress trade-offs, MCP exposure for WooCommerce, Core Web Vitals on large catalogs, Astro versus Next.js for content-heavy stacks. They sit between SEO top-of-funnel and sales-ready intent.

We track them because AI answers often replace the first three blue links a technical evaluator would have opened. If you are absent here, you are not in the shortlist conversation even when your identity prompt ranks well.

#Transactional WooCommerce prompts

This is where budgets live: hire WooCommerce developer, WooCommerce ERP integration partner, fix slow checkout on a large Woo store, agency for Woo plus AI automation. The Geoboard baseline on 2026-06-11 showed zero presence for wppoland.com across this family.

That zero is the KPI the 90-day series watches. Not because on-page copy is finished, it largely is, but because models cite what the open web corroborates. Directories and broad “AI SEO” agencies win here on third-party mentions, not on cleaner FAQ schema.

#AI-implementation prompts

Parallel to WooCommerce: WordPress AI integration, MCP server for store data, agent-ready product feeds, governance for AI plugins. Same baseline result: zero at 2026-06-11. These prompts matter for enterprise procurement packages that now bundle “AI readiness” with platform work.

If your vendor deck shows one aggregate score across these rows, ask for the table instead.

#Methodology

We use three layers. None is perfect. Together they are auditable, which is the bar procurement should enforce.

#Geoboard batch baselines

Geoboard gives repeatable multi-model runs on a declared prompt set. Our frozen baseline is dated 2026-06-11. On the anchor identity prompt, wppoland.com ranked first in five of six models. The same run showed the transactional split above and surfaced engine skew: Perplexity strongest, ChatGPT weakest, including a 0/8 blind spot on the ChatGPT slice of that prompt set.

We re-run Geoboard on a six-to-eight-week cadence during the 90-day window, not daily. Daily multi-model polling mostly measures UI noise. The next scheduled batch re-run is the earliest point we will call a “trend”, not this article.

#Weekly manual spot checks

Every week we re-ask the same fixed prompts in consumer interfaces: ChatGPT, Perplexity, Copilot, and Claude where applicable. We log four fields: date, engine, prompt family, and outcome (brand mention, URL cited, competitor domains cited, or none).

Weekly checks catch product changes fast. They are not a substitute for batch tooling. They are the early-warning layer when a model adds or removes browsing.

#A result record that another person can challenge

The useful unit in our log is not a screenshot. It is one dated observation tied to an exact prompt and product surface. For each run we preserve the prompt wording, assistant, visible browsing state, outcome category and domains shown as sources. We also distinguish a plain-text brand mention from a clickable citation to wppoland.com. Those events have different commercial value and should never be collapsed into the same green tick.

We keep the prompt wording fixed inside the comparison set because even a helpful rewrite can change the intent. A phrase such as “best agency” invites a list; “who can integrate WooCommerce with an ERP” asks for capability; “Polish WordPress agency for an overseas client” tests identity and geography. If a prompt must change because a product interface changes, it starts a new series rather than silently continuing the old one.

The same principle applies to personalisation. A signed-in assistant may use conversation history, location or prior browsing. Manual checks therefore remain observations of a declared surface, not universal rankings. We do not claim that every buyer will receive the same answer. The question is narrower: under a repeatable setup, is the brand retrieved, cited and associated with the intended service more often than at baseline?

#What counts as movement, and what does not

A single new mention is a signal for inspection, not proof of durable growth. We first check whether the assistant cited our own page, repeated a third-party description or merely produced the brand from its model memory. We then look for the same association in another run or engine. This prevents an isolated answer from becoming a sales claim.

The strongest commercially relevant movement would be a cited URL on a transactional prompt, followed by repeated retrieval across dates. An unlinked identity mention is weaker. A citation on an informational query can still matter because it places expertise in the buyer’s research path, but it does not demonstrate hire intent. Conversely, holding the identity position while transactional prompts remain empty is not “overall improvement”. It is stability in one family and no measured progress in another.

#Honest limitations we publish up front

API proxies are not consumer products. Our Q2 work showed proxy runs that returned “undetermined” for every query while consumer monitoring showed a real split. We do not mix those numbers in this series.

Weekly notes are not ninety comparable chart points yet. Today is 2026-07-10. We have a baseline plus the discipline going forward. We will not fabricate intermediate weekly aggregates for content marketing.

Engine heterogeneity is real. Perplexity leaning on live search and ChatGPT showing a blind spot on the same prompt set is expected, not a bug in our site. Reporting one average hides the channel you are actually losing.

Rendering matters. Andre Alpar’s June 2026 live test published via CitationOne found six of seven Western assistants returned content from raw HTML only, not executed JavaScript. That aligns with our stack choice and with our article on Western AI assistants not rendering JavaScript. If your “GEO fix” depends on client-side hydration, measurement will keep reporting zero no matter how polished the copy.

#What we found

Frame this section correctly: it is baseline plus interpretation, the starting line of the 90-day series, not the finish.

#Identity is defensible

On 2026-06-11, for Polish agency for foreign WordPress clients, we ranked first in five of six models. That matches deliberate positioning: Poland-based, foreign-client focus, WordPress as the core stack. A procurement evaluator typing that exact narrowing question gets an answer that includes us. That is useful and limited.

#Transactional citations are absent

On WooCommerce hire and AI-implementation prompts, the baseline presence was zero. Competitors surfacing in manual checks were often generalist SEO or SEM shops advertising “AI SEO”, not Woo specialists. That tells a procurement story: the model’s association graph for money queries does not yet include us, regardless of on-page schema depth.

#ChatGPT blind spot versus Perplexity strength

Within the baseline prompt set, ChatGPT showed zero presence across eight tracked prompts in the blind-spot slice. Perplexity was the strongest engine in the same run. For vendor planning, that means a ChatGPT-only check systematically under-reports a brand that Perplexity might already cite. Any monitoring scope that excludes Perplexity is incomplete for B2B research behaviour in 2026. We unpack the mechanism and fix paths in why Perplexity cites your brand but ChatGPT does not.

#Rendering is a gate, not a tie-breaker

The CitationOne/Alpar June 2026 experiment is independent of our Geoboard run, but it explains a mechanical ceiling: if assistants read raw HTML, your citations depend on what is in the initial response bytes. We ship server-rendered HTML and structured data for that reason. It is necessary. It did not, by itself, unlock transactional citations in the baseline.

#How this differs from our Q2 report

The Q2 measurement article documented three snapshots and proxy versus consumer disagreement. This series holds prompts constant for ninety days to see whether off-page authority work moves the transactional zero. We are not re-litigating proxy error modes here; we are operating one combined instrument going forward.

#What this means for WordPress site owners

#For procurement and vendor governance

Treat AI visibility like any other vendor KPI: require method, prompt list, engines, and dates. Reject composite scores that blend identity wins with transactional losses. Ask where the baseline was taken and whether weekly numbers come from consumer UIs or APIs.

When evaluating a WordPress agency for GEO, ask for a split table like the one above, not a single “citation score”. Our baseline says we are credible on identity and not yet credible on Woo hire intent. That is what honest first-party reporting looks like.

#For site and platform teams

On-page work still matters for informational prompts and for being machine-readable. Server-rendered HTML, stable schema, speakable selectors, and fact-dense llmCard blocks are the floor. The AI and LLM visibility playbook orders those levers the way we implement them in client programmes.

Transactional gaps, however, rarely close with another FAQ block. They close when independent surfaces corroborate you: technical answers on Stack Overflow, open-source artifacts, directory profiles with verifiable reviews, podcast or guest appearances. That is the same off-page plan we document internally and execute outside the repo.

#For marketing and SEO leads

Do not publish weekly citation charts unless you can publish the prompt list and engine log alongside them. Smooth curves without method are a reputational risk in 2026, especially after public experiments showing models hallucinating brands that never existed.

Prefer quarterly or window-based readouts. We chose ninety days because off-page placements need weeks to be crawled, associated, and retrieved. Identity can move in a single crawl cycle; transactional associations do not.

#For commercial teams deciding what to do next

The measurement should change a decision, otherwise it is reporting theatre. A missing informational citation suggests that the page may not answer the question clearly enough, expose the answer in initial HTML or provide a source worth citing. A missing transactional association points towards a different backlog: corroborating case studies, credible third-party profiles, specialist contributions and clearer links between the company entity and the service.

It is equally important to avoid attributing every enquiry to GEO. A prospect may see an AI answer, return through branded search and submit a form days later. Citation monitoring shows retrieval behaviour; analytics and lead qualification show commercial impact. We would examine the landing page, referral information where available, assisted journeys and the prospect’s own account of how they found the company. None of those fields alone establishes causation.

For a budget owner, the sensible 90-day question is therefore not “did our AI score rise?” It is “which query family moved, what evidence changed before it moved, and did that movement place us in a relevant buying journey?” That framing makes a zero useful too. It can stop spend on repetitive on-page decoration when the missing ingredient is independent authority.

For the full cadence table, metrics list, and procurement checklist, see our dedicated guide on AI-citation monitoring: what to track and how often.

#How we monitor going forward

This is the operating cadence for the series ending 2026-10-08 (ninety days from 2026-07-10).

Weekly: same prompts, consumer UIs, four-column log (engine, date, outcome, cited domains). No public chart until the window closes.

Every six to eight weeks: Geoboard batch re-run on the full prompt set. Primary success signal is first non-zero presence on transactional WooCommerce or AI-implementation prompts, today still zero at baseline.

Ad hoc: manual checks within 48 hours when an assistant ships a browsing or UI change.

Parallel: brand-radar style mention tracking for competitor citation domains, mapped to the off-page authority plan. Geoboard tells you if you appeared; radar helps explain who else models cite instead.

Publication rule: one honest readout after day ninety on the same prompts. If a mid-window anomaly is material, we will note it with method attached, not imply a trend from a single week.

If you want the programme version of this instrumentation applied to your property, that is what GEO and LLMO optimization delivers: baseline, prompt families, engine split, and work tied to the gap type. If you want the strategy map first, start with the AI and LLM visibility playbook.

To make an initial assessment concrete, send us a written brief rather than a request for an undefined “AI visibility score”. Include the domain, countries and languages that matter, three to ten buyer questions you would genuinely want an assistant to answer, the services linked to revenue, and any monitoring already in place. Add known competitors and state whether you need a diagnostic baseline, an implementation programme or independent validation of an existing report. We can then map the brief to a measurable scope without pretending that one generic prompt set represents your market.

#Closing note

We are not asking you to trust a vanity metric. We are asking you to demand the same traceability from us that you should demand from any vendor pitching “AI search dominance”. The Geoboard baseline on 2026-06-11 is the numbered starting line: strong on identity, absent on transactional Woo and AI implementation, skewed by engine, gated by HTML rendering. The next meaningful update comes when the ninety-day window completes, measured the same way, prompts unchanged.

Until then, the useful takeaway is smaller and more valuable: measure in families, publish the method, and do not confuse a positioning win with procurement-ready visibility.