1.Executive summary

Checks summary

2.AI crawler access

Content is available without JavaScript. This matters: models do not execute scripts. On a product page the raw HTML holds 24,591 characters of text against 18,044 in the rendered DOM — meaning more is visible without JS, not less. The source code contains the price, full specifications, a comparison table and ProductGroup markup with 15 product variants.

What is closed and needs action

The knowledge base is unavailable to anyone. support.evapolar.com (Zendesk Help Center) returns 403 to everyone — including a regular browser and Googlebot. This is a Cloudflare challenge at the Zendesk configuration level, not anti-AI protection. The consequence: maintenance instructions, answers about "it isn't cooling", warranty terms — all of it sits outside Google's index and outside the models. And service questions are exactly what people most often ask assistants.

The challenge behaves inconsistently between checks, so client-side diagnostics should start with the WAF settings in the Zendesk account.

agents.md and llms.txt exist but are empty of meaning. Shopify serves them automatically — they advertise the platform and say nothing about Evapolar: no product line, no core differentiator ("evaporative cooler ≠ air conditioner"), no FAQ. The files are editable. This is the cheapest lever in the entire audit: a single file read specifically by AI agents.

Common Crawl could not be checked — the API returned 503 across more than 30 attempts. The initial "0 URLs across four indexes" result is an artifact of that failure, not evidence of absence; it was excluded from the report.

3.Technical audit

Non-canonical URLs: 48% of impressions off target

The main sources: legacy /blog/ addresses instead of /blogs/blog/ — 32.6% of impressions, and www instead of the main domain — 15.0%.

The redirect is only partially rolled out. A check of 40 random real articles: 26 return 301, 14 return 404. The rule works on popular pages but not on the tail. What is needed is a blanket pattern redirect across the whole prefix, including /blog/es/, not a page-by-page list.

Separately, 40 AMP addresses (23,977 impressions, one at position 1.6) and 130 URLs of the form /blog/content/images/... remain unhandled.

The removed French locale returns 404

/fr/ was deleted without redirects to equivalents. The pages remain in the index: 20 of the top 25 fr addresses return 404, for a total of 16,665 impressions. One URL sits at position 1.1 with 9,843 impressions and serves an error.

The root addresses (/fr/, /fr/shop) already redirect — it is the articles and products that are unhandled.

Infinite URL space

?page=5000 on tag listings returns 200 and a self-canonical. With 342 listings the URL space is literally infinite — this broke the full site crawl twice. The index is protected by a noindex directive, but crawl budget is spent for nothing, including by the models' crawlers.

Canonical conflict

111 pages serve two mutually exclusive rel=canonical tags at once — one to self and one to the blog root. Faced with a conflict, Google discards both signals.

A template bug that breaks two things at once

The theme template drops a slash when concatenating addresses. Two symptoms:

• "logo": "https:files/logotype_gray.svg" in the organization markup — on all 1,068 pages, and the address returns 404;
• author.url of the form /depages/eugenedubovoy and /espages/eugenedubovoy — 404 on 207–215 articles, while the correct /de/pages/... and /es/pages/... work fine.

Fixed with a single edit. This is the first candidate for the roadmap.

Core Web Vitals

For real users (CrUX, 07/18–08/14/2026, mobile), everything is in the green:

The only deviation is INP 216 ms on product pages, above the threshold. The cause is visible in the lab test: 953 KB of unused JavaScript and 6.3 s of main-thread work.

A lab Performance score of 38–43 alongside green field data is normal: Lighthouse emulates a slow device. There are no penalties for this score and no need to optimize for it. But it does show an absence of headroom: the site passes thanks to a good audience, and adding one more script would push product pages into the red first.

📊 broken-links · 📊 redirects · 📊 duplicate-titles

4.Readiness for AI answers

The company does not exist in machine-readable form

Organization markup is the only way to tell a model what company stands behind the content. On evapolar.com it is broken identically on all 1,068 pages:

• there is no @id — the entity has no identifier, and there is nothing to link the articles to;
• logo points to a 404;
• sameAs holds five empty strings out of eight, and the Facebook link returns 400;
• there is no address, legalName (the legal entity is CoolTrade Inc.), foundingDate, founder or contactPoint.

The defect is systemic — it lives in a shared theme file. That makes the fix cheaper (one edit repairs three locales) and explains the "20% in AI Overviews against 5% in ChatGPT" gap.

The proof of authority exists but is not published

The company's open cloud storage holds a ready-made press kit: logos from Wired, Forbes, TechCrunch, Mashable, PCMag, BuzzFeed, Business Insider, Engadget and a press release with verifiable facts — three international awards in 2017, a finalist place at the IOT/WT Innovation World Cup, 30,000 devices in 125 countries.

On the site there is not a single mention. The pages /pages/press, /pages/awards, /pages/team and /pages/reviews return 404. The press release points to evapolar.com/press — also a 404.

Amazon reviews marked up as first-party — a penalty risk

On 15 product addresses (all three locales), reviews sourced from Amazon are marked up as reviews of the product on this site. The section itself is labeled "Verified reviews from Amazon customers" — that is, the site acknowledges the third-party origin and serves them as its own anyway.

All 10 reviews carry a rating of 5, the publication date is missing, and a service attribute is duplicated 60 times instead of 10, spawning phantom product entities.

This is a direct violation of the Google Review snippet policy. The validator stays silent about it: Screaming Frog reports 0 errors across all 1,068 pages — what is broken is the policy, not the syntax.

This is the one finding that should be fixed before all the others.

The blog does not lead to the store

Of 415 articles, only 14 (3.4%) contain a link to a product or collection in the body text. Formally links are everywhere, but by position they are 4,800 navigational and 400 footer links — that is, the menu, identical across the whole site.

Articles with no commercial path collect ≈39,900 visits per month and generate $0.

A telling example: /blogs/blog/air-conditioner-does-not-keep-rv-cool — 20,454 visits per month, while the /collections/for-rv-camper collection exists right alongside it and the article does not link to it.

The blog's own quality is high: extractability 97, E-E-A-T 88 — higher than on the commercial pages. The issue is not the articles, it is that there is nowhere to go from them.

Product pages say nothing about the main advantage

On /products/evachill, /products/evalight, /products/evasmart and the homepage there are zero mentions of "without a window", "ventless", "without hose", "exhaust" or "freon". The English and German versions are the same.

The word "window" appears 4 times on the evaCHILL page — all inside third-party reviews, and in a sense that undermines the positioning: "if you do not ventilate your room, eventually the room will get too humid".

The right wording does exist — but only on collection pages, which have 8× fewer impressions and 117 inbound links against 1,774 for the general catalog. None of the six articles ranking for "without a window" queries links to the relevant collection.

The "without a window" cluster: 413 queries, 330,910 search volume, delivering 377 clicks per year.

Additionally on the money pages: the target query is missing from the H1 on all 10 pages, there is not a single table (comparisons are laid out as flat blocks), FAQPage is not marked up anywhere despite questions being present on 8 pages, and 60–79% of a product page's text is third-party reviews quoting the outdated $79 price against the current $99.

Content freshness is fictitious

415 dated pages carry just 19 unique modification-date values; seven values cover 326 pages. 227 pages older than two years are flagged as updated within the last 180 days.

Recalculated by publication date: 191 pages are genuinely outdated, including 10 comparisons up to 4.3 years old. Another 217 pages have no date at all — all product, category and landing pages.

Duplicates

43 pairs of fully identical text, 86 addresses — 16% of the indexable corpus. Both copies in every pair are indexable and self-referencing, with not a single noindex directive.

📊 pages · 📊 insights · 📊 money-pages · 📊 ctr-gap · 📊 duplicates

5.Visibility in model answers

The paradox: cited first, recommended last

443 answers mention competitors but not Evapolar. The combined volume of those prompts is 1,496,500.

The gap between first place by citations and fifth by mentions is the measurable expression of the problem: the content is used as a source of facts, but not as a brand recommendation.

Traffic from models is the best channel in revenue terms

GA4 data for the clean window (April–August 2026):

ChatGPT converts 7.4× better than organic, and ARPU is 6.8× higher. The channel grew 5.9× over six months. ChatGPT accounts for 85% of the channel's traffic and 98% of its revenue.

Mobile delivers 82% of channel revenue on 69% of its traffic. The best ARPU is Switzerland ($5.74).

Urgent: six months of revenue data lost

From October 2025 through March 2026, e-commerce attribution was not working: 120 thousand sessions, zero transactions and zero revenue across every channel at once. This is not a drop in sales but a failure in event delivery.

All the figures above are calculated on the clean window. Before making any channel decisions, confirm that the purchase event is arriving consistently.

Separately: GA4's native "AI Assistant" channel sees 1,431 sessions against 2,577 under the extended rule — a 44% understatement of the channel.

6.Semantics & content plan

The plan: 64 clusters, 41 new pages, 23 updates to existing ones; 42 clusters matter specifically for visibility in models.

A telling US comparison: Evapolar ranks for 1,201 queries but only 100 in the top 10. Portacool has 261 in the top 10 on a smaller overall footprint.

An anomaly worth a quick check: best portable air cooler — 40,500 searches per month, difficulty 15, and the site sits at position 22.6 despite having a dedicated page.

Seasonality is severe: July delivers 10.6× more clicks than October. Content for the 2027 season must be published in February–April — indexing takes 2–4 months, and June is already too late.

📊 keywords-gap · 📊 page-plan · 📊 content-gap

7.External placements

The main opportunity is not a link, it is a line in a roundup

bobvila.com runs a review titled "The Best Evaporative Air Coolers of 2026" and recommends Uthfy, Arctic Air, Portacool, Dreo, Hessaire, Soleey — Evapolar is absent, even though four competitors have links there. sylvane.com has its own roundup, "Best Evaporative Swamp Coolers 2026", also without Evapolar.

These are exactly the pages models cite when asked for the best personal evaporative cooler. Being absent from them subtracts more from visibility than any link-buying program adds.

The winners on the missed prompts are review sites (rtings 379 citations, Consumer Reports 333, Lowe's 583), not competitor sites. The lever here is external placements, not on-site fixes.

What not to buy

Of 1,000 donors, 504 are classified as "do not replicate" — networks on free hosting and PBNs. Some of them pass the formal spam filter (score 27 against a threshold of 30), so the list was assembled manually from host patterns.

Specialized climate-equipment directories open to paid placement do not exist — the vertical monetizes through reviews. The entire HVAC segment works through pitching only.

📊 authority-placements · 📊 inventory-buylist · 📊 pitch-targets · 📊 do-not-replicate

8.Roadmap

A.Methodology & limitations

Appendix A. Methodology and limitations

Sources: Screaming Frog crawl on a dedicated droplet (1,068 HTML pages, three locales), GA4 Data API (property 272829378), Search Console API (12 months), CrUX and PageSpeed Insights, DataForSEO (Backlinks, Labs, LLM Mentions), Ahrefs Brand Radar, Google Ads Keyword Planner, live requests from 14 User-Agents × 7 pages, manual JSON-LD collection.

What could not be checked

• Common Crawl — the API was unavailable (503, >30 attempts). The initial "0 URLs" result is an artifact of the failure and is excluded from the conclusions.
• Prompt checks in model interfaces — the browser automation tool was unavailable. Compensated with samples of 100 real model answers via API, which is stronger evidence than manual checks.
• The Core Web Vitals report from Search Console — requires extended access permissions; the CrUX data substitutes for it.
• Server logs were not provided — crawler activity was estimated from indirect signals.

Caveats on the figures

• Revenue for October 2025 – March 2026 is unreliable (attribution failure). All ARPU and conversion figures are calculated on the 04/01–08/15/2026 window, which falls within the seasonal peak.
• broken-links was cleaned before delivery: the raw export contained 10,415 rows, but 6,863 were repeats of 22 footer links and 3,395 were artifacts of crawler rate limiting. The tab contains 59 unique broken addresses, of which 3 are internal.
• Reddit data is estimated — access to the platform was blocked.
• Absolute DataForSEO and Ahrefs figures were not mixed — the services use different indexes (759 answers against 73 mentions for one and the same brand). Metrics are given with their source stated.
• Competitor positions were taken for the US; the plans for Germany and Spain are built on Keyword Planner and Search Console.
• German locale: the translations exist (verified with live requests), but most have no internal links — they are reachable only through pagination, so the crawler saw 44 of ~192 articles. This is a finding about internal linking, not about missing content.
• The status of the legacy redirects was clarified by a separate check: on a sample of 40 random articles, 26 return 301 and 14 return 404. The discrepancy with earlier estimates is explained by those being made on top pages, where the rule does work.
• The markup was not run through Google's validator — the defects were identified by code analysis; the markup is syntactically valid, but the semantics and the policy are violated.

Appendix B. Section reports

Data: client spreadsheet — 16 tabs.

Appendix C. Ready-made assets

• Organization JSON-LD with @id, a working logo, legalName, foundingDate, founder and an extended sameAs — E-E-A-T section
• Person JSON-LD as a standalone entity with correct profiles — same section
• BreadcrumbList, FAQPage, CollectionPage and Article JSON-LD with an end-to-end @id graph — structured data section
• Copy for the "Works without a window" block and rewritten H1s for 6 money pages — on-page section
• A breakdown of the 43 duplicate pairs with the recommended 301 direction — duplicates tab
• A list of 38 articles with the target collection for a contextual link — insights tab