Search visibility, AI discovery, paid efficiency, an external read on what's working, what isn't, and what to act on this cycle. Scoped to US users on roomsketcher.com.
Every figure below is filtered to United States users on roomsketcher.com. Session and click metrics are reported as daily averages so windows of different length stay comparable. The Δ column compares Current against Baseline (the immediately prior 14 days); the Last Reported column is what Issue #7 showed, included here only because five weeks passed between calls.
| KPI | Last Reported Jun 15 – Jun 28 | Baseline Jul 6 – Jul 19 | Current Jul 20 – Aug 2 | Δ vs Baseline | Status |
|---|---|---|---|---|---|
| Website Traffic (US) — daily averages | |||||
| Total Website Sessions | 4,218/day | 4,459/day | 4,350/day | -2.4% | |
| Direct Sessions | 2,121/day | 2,243/day | 2,231/day | -0.5% | |
| Organic Search Sessions | 1,212/day | 1,309/day | 1,102/day | -15.8% | |
| Paid Search Sessions (Google) | 328/day | 348/day | 460/day | +32.0% | |
| Paid Other Sessions (Microsoft) | 228/day | 267/day | 232/day | -12.9% | |
| Referral Sessions | 69/day | 78/day | 74/day | -4.0% | |
| Email Sessions | 39/day | 41/day | 38/day | -6.7% | |
| Organic Social Sessions | 29/day | 37/day | 33/day | -10.4% | |
| Bing & Alt-Engine Organic (US) — daily averages | |||||
| Bing Organic Sessions | 73.1/day | 84.1/day | 77.9/day | -7.5% | |
| DuckDuckGo Sessions | 79.6/day | 92.1/day | 82.9/day | -10.0% | |
| Yahoo Sessions | 26.3/day | 27.5/day | 27.3/day | -0.8% | |
| AI Referral Traffic (US) — daily averages | |||||
| Total AI Referral Sessions | 34.4/day | 29.7/day | 25.2/day | -15.1% | |
| ChatGPT | 28.9/day | 23.1/day | 19.7/day | -14.8% | |
| Gemini | 2.1/day | 2.2/day | 2.0/day | -9.7% | |
| Perplexity | 2.1/day | 3.1/day | 2.1/day | -34.1% | |
| Copilot | 0.8/day | 0.9/day | 0.6/day | -25.0% | |
| Claude | 0.6/day | 0.4/day | 0.8/day | +120.0% | |
| Search Console (US) | |||||
| Clicks Current window has 13 days of GSC data (normal lag) | 1,146/day | 1,198/day | 1,001/day | -16.5% | |
| Impressions | 107,516/day | 100,934/day | 94,857/day | -6.0% | |
| Non-Brand Clicks branded queries excluded | — | 1,109/day | 918/day | -17.2% | |
| Non-Brand CTR | — | 1.103% | 0.972% | -0.13pp | |
| Average CTR | 1.066% | 1.187% | 1.055% | -0.13pp | |
| Average Position | 10.02 | 9.91 | 10.26 | +3.5% | |
| Key Head-Term Positions (US) — lower is better | |||||
| “floor plan software” | — | 9.6 | 8.8 | -8.3% | |
| “floorplan software” | — | 7.0 | 6.9 | -1.4% | |
| “floor plan app” | — | 10.0 | 8.9 | -11.0% | |
| “free floor plan software” | — | 10.1 | 5.5 | -45.5% | |
| “floor plan” | — | 12.5 | 7.4 | -40.8% | |
| “room planner” | — | 29.9 | 20.7 | -30.8% | |
| “floor plan creator” | — | 18.7 | 21.0 | +12.3% | |
| “floor plan maker” | — | 24.8 | 24.1 | -2.8% | |
| Template Gallery (/floor-plan-gallery/) — scale corrected this issue | |||||
| Gallery Clicks | — | 90.6/day | 90.6/day | +0.0% | |
| Gallery Impressions | — | 12,642/day | 12,223/day | -3.3% | |
| Gallery Pages Receiving Impressions | — | 833 | 832 | -0.1% | |
| Gallery Pages Receiving Clicks | — | 157 | 166 | +5.7% | |
| Google Ads — Spend & Efficiency (US, NOK converted at 0.091) | |||||
| Spend | 1,111/day | 985/day | 1,602/day | +62.6% | |
| Clicks | 595/day | 688/day | 926/day | +34.4% | |
| CTR | 8.27% | 9.81% | 9.84% | +0.03pp | |
| Total Conversions (all actions) | 97.5/day | 125.7/day | 150.4/day | +19.6% | |
| Blended Cost per Conversion | 11.40 | 7.84 | 10.65 | +35.9% | |
| Google Ads — Free Signups (US) | |||||
| Free Account (Personal) | 79.4/day | 109.1/day | 130.9/day | +20.0% | |
| Free Account (Business) | 8.6/day | 9.1/day | 10.9/day | +20.5% | |
| Google Ads — Paid Conversions (US) | |||||
| New Subscription | 7.26/day | 6.11/day | 7.11/day | +16.4% | |
| Cost per New Subscription | 153 | 161 | 225 | +39.7% | |
| Premium Project | 1.61/day | 0.89/day | 1.07/day | +20.0% | |
| Ready Made Purchase | 0.64/day | 0.57/day | 0.36/day | -37.5% | |
| Microsoft (Bing) Ads (US) — NOK converted at 0.091 | |||||
| Spend | — | 569/day | 382/day | -32.9% | |
| Clicks | — | 518/day | 432/day | -16.5% | |
| Total Conversions | — | 44.4/day | 33.2/day | -25.2% | |
| Blended Cost per Conversion | — | 12.82 | 11.50 | -10.3% | |
| New Subscription | — | 3.01/day | 1.64/day | -45.6% | |
| Cost per New Subscription | — | 189 | 234 | +23.4% | |
| Combined Paid (Google + Microsoft, US, USD) | |||||
| Combined ROAS conversion value ÷ spend — Q3 KR target > 1.0 | — | 0.81 | 0.60 | -25.9% | |
| Google Ads ROAS (US) | — | 0.87 | 0.61 | -29.9% | |
| Google Ads ROAS incl. renewals | — | 1.01 | 0.73 | -27.7% | |
| Microsoft Ads ROAS (US) | — | 0.72 | 0.57 | -20.8% | |
| Spend-weighted Search Impr. Share | — | 32.7% | 31.8% | -0.9pp | |
| Combined Spend | — | 1,555/day | 1,984/day | +27.6% | |
| Combined Conversions | — | 170.1/day | 183.6/day | +7.9% | |
| Combined Cost per Conversion | — | 9.14 | 10.81 | +18.3% | |
| Combined New Subscriptions | — | 9.12/day | 8.75/day | -4.1% | |
| Combined Cost per New Subscription | — | 171 | 227 | +33.1% | |
| GA4 Conversion Events (All Sources, US) | |||||
| Sign-Up Rate sign_up ÷ sessions, US | — | 7.04% | 7.00% | -0.6% | |
| sign_up | 274.1/day | 314.1/day | 304.4/day | -3.1% | |
| begin_checkout | 53.7/day | 54.8/day | 55.5/day | +1.3% | |
| new_subscription | 17.9/day | 17.3/day | 16.6/day | -4.1% | |
| purchase | 33.6/day | 32.3/day | 30.9/day | -4.4% | |
| Site Performance (local Lighthouse — see Section 7, not PSI-comparable) | |||||
| Mobile Performance (median of 4) | — | — | 84 | — | |
| Mobile LCP (median of 4) | — | — | 4.01s | — | |
| Desktop Performance (avg of 2) | — | — | 94 | — | |
| Desktop LCP (avg of 2) | — | — | 1.48s | — | |
| Accessibility | — | — | 97 | — | |
| SEO | — | — | 100 | — | |
sign_up is down only 3.1% and begin_checkout is up 1.3%, against a 16% organic session decline and a 28% paid spend increase. Demand is roughly intact; the efficiency of buying it is what changed.Channel mix for the current window, July 20 – August 2, 2026. 60,898 total US sessions. Direct remains over half of all traffic — a persistent measurement question as much as a channel, addressed below.
Reading the current window against the baseline alone understates what is happening. Weekly US organic sessions since the tracking rebuild show a steady, uninterrupted slide — and the final full week of this window is the lowest on record since measurement became trustworthy.
Weekly US organic sessions have fallen from ~11,500 in March to 7,318 in the week of July 26, a 37% decline over five months with no single-week cliff. That is the symptom. The cause is narrower and more actionable than the headline suggests.
First, what kind of loss it is. Comparing March 8–April 4 against June 28–July 25: impressions fell 26.6% (133,358→97,856/day) while CTR actually improved 6.9%. Decomposed, that is −374 clicks/day from lost impressions and +72/day recovered by better click-through. The site is being shown less, not clicked less. It is not de-indexing either — 289 pages dropped out of results (0.4% of impressions) and 235 new ones replaced them, leaving the footprint stable at ~1,600 pages.
Second, where the impressions went. They are not spread evenly. Nine pages account for 55% of the entire decline, and every one of them lost rank:
| Page | Impr/day Mar | Impr/day Jul | Δ | Position |
|---|---|---|---|---|
| /house-plans/ | 3,933 | 1,230 | −69% | 11.8 → 27.7 |
| /kitchen-planner/ | 3,054 | 605 | −80% | 12.1 → 34.2 |
| /blog/how-to-measure-floor-area/ | 1,322 | 265 | −80% | 2.6 → 5.1 |
| /features/roomsketcher-app/ | 1,666 | 760 | −54% | 8.1 → 12.9 |
| /features/draw-floor-plans/ | 1,597 | 706 | −56% | 11.4 → 20.9 |
| /blog/how-to-read-floor-plans-9-easy-steps/ | 798 | 92 | −88% | 3.6 → 6.4 |
| /download-app/ | 637 | 194 | −70% | 2.0 → 5.4 |
| /subscriptions/free/ | 549 | 146 | −73% | 2.3 → 3.4 |
| /home-improvement/home-remodeling-software/ | 506 | 179 | −65% | 14.2 → 27.0 |
| These nine pages | 14,062 | 4,177 | −70% | 55% of the decline |
/house-plans/, /kitchen-planner/ and /home-remodeling-software/ all went from page 1–2 to roughly position 28–34. A 20-position fall is structural, not drift. The other six slipped from strong positions; note that /blog/how-to-read-floor-plans-9-easy-steps/ lost 88% of its impressions on only a 2.8-position slip, which is the difference between sitting in the answer-box zone and sitting below it./blog/how-to-draw-a-blueprint/ went 10.1→3.9 with impressions down 27%. That residual is answer-engine compression and seasonality, and no amount of page-level work reverses it.Why this matters for the plan: a uniform algorithmic squeeze offers nothing to act on. Nine named URLs losing 5–22 positions each is a work list. Q3 Objective 2's Key Page UX Update (10 high-traffic feature/landing pages, owner Trude, due Sep) is aimed at exactly this shape of page — /features/draw-floor-plans/, /features/roomsketcher-app/ and /kitchen-planner/ all qualify. That target list was drawn up before this analysis existed and should be checked against this table.
Paid Search rose 32% in the same window organic fell 16%, which invites the obvious question: are we simply buying traffic we used to get free? Tested four ways, the answer is no — with one small, real exception.
The harder read: a straight trade would at least be neutral — total sessions flat, cost shifted. That is not what happened. Total US sessions fell 2.4% (4,459→4,350/day) while combined paid spend rose $429/day. Paid did not take organic's traffic; it partially and expensively backfilled an organic decline that has its own separate cause. Two independent problems, not one trade-off.
For the portion of the decline that page-level work cannot explain, there is no named event to blame or wait out. Google confirmed on July 9 that smaller core updates now roll continuously without public announcement, and third-party volatility trackers recorded sustained movement through June and July with no official label. There was no announced core update in July 2026. The practical consequence: no rollout end-date, no recovery event to expect, and CTR-at-stable-position becomes its own metric to watch. It is also the part of the problem that argues for answer-surface presence over incremental blue-link optimisation.
Direct is 51.3% of US sessions and converts at 9.5% to sign_up — roughly 4.7× the Organic Search rate (2.0%). A cold-traffic channel does not behave that way. It is overwhelmingly returning users and app-adjacent sessions where GA4 kept the user but lost the source. But this cycle's daily data shows it is not purely a retention signal either.
The channel chart shows Direct peaking in the week of July 20 and dropping sharply the following week. The daily data explains it, and the explanation matters for how paid budget decisions get evaluated.
| Date | Direct sessions | Search-NORAM spend | What happened |
|---|---|---|---|
| Jul 13–17 (Mon–Fri) | 2,378/day | — | Normal baseline |
| Jul 21–23 | 2,965 / 3,566 / 3,112 | 16,817 / 16,822 / 16,847 NOK | Direct spikes on exactly the three days paid spend was pinned at its cap |
| Jul 27–31 (Mon–Fri) | 1,969/day | 593–3,082 NOK | Budget cut ~85%; Direct falls with it |
AI referral traffic — sessions arriving from an AI assistant's answer — fell 15.1% against the baseline and 26.8% against the last reported period. The LLM Visibility Score slipped on both anchor queries over the same stretch, but by less — and RoomSketcher is still the outright #1 recommendation on three of six engines, including the one that sends 78% of the traffic.
| Engine | Last Reported Jun 15 – 28 | Baseline Jul 6 – 19 | Current Jul 20 – Aug 2 | Δ vs Baseline |
|---|---|---|---|---|
| ChatGPT (chatgpt.com) | 404 | 324 | 276 | −14.8% |
| Perplexity (both sources) | 30 | 44 | 29 | −34.1% |
| Gemini (gemini.google.com) | 29 | 31 | 28 | −9.7% |
| Claude (claude.ai) | 8 | 5 | 11 | +120.0% |
| Copilot (copilot.com) | 11 | 12 | 9 | −25.0% |
| Total AI Referral Sessions | 482 | 416 | 353 | −15.1% |
perplexity.ai and perplexity. bing and cn.bing.com are counted as Bing organic, not AI.AI referral sessions dropped 15.1% while the LLM Visibility Score slipped only 8.3 and 4.2 points on its two anchors — and RoomSketcher is still the outright #1 recommendation on ChatGPT, Perplexity and Copilot. The traffic is falling faster than the ranking, which points at in-line answering absorbing clicks rather than at losing the recommendation.
The GA4 “AI Assistant” channel produced 16 sign_up events on 336 sessions — a 4.8% signup rate, against Organic Search's 2.0% (310 signups on 15,430 sessions). AI-referred visitors arrive pre-qualified: the assistant has already answered "which tool should I use," so the visit is a decision confirmation rather than a comparison. The volume is small, but per-session it is the highest-intent non-branded traffic on the site.
ChatGPT alone is 78% of AI referral traffic, so total AI volume is effectively a ChatGPT metric — concentration risk worth naming. Claude more than doubled (5→11 sessions) but off a base small enough that two weeks of noise could explain it; worth watching rather than acting on. Perplexity's 34% decline is the sharpest fall and is the one engine where the June reading showed a genuine ranking weakness on the “software” phrasing.
Two fixed anchor queries run across six engines on a New York VPN, in fresh/temporary chats so conversation memory cannot bias the answer. Each cell is scored 0–1: 1.0 = definite #1 / best-overall, 0.75 = wins a named sub-category while another tool or nothing holds the overall slot, 0.5 = listed among the recommendations, 0.25 = mentioned or cited but not recommended, 0 = absent. This is the fourth reading; the first was June 1.
| Engine | “floorplan software” Jun 29 | “floorplan software” Aug 3 | “floor plan creator” Jun 29 | “floor plan creator” Aug 3 | What the answer said this reading |
|---|---|---|---|---|---|
| ChatGPT | 1.0 | 1.0 | 1.0 | 1.0 | “Best overall: RoomSketcher” in both. Creator answer ranks it 🏆 #1, “Overall best.” |
| Perplexity | 1.0 | 1.0 | 1.0 | 1.0 | “Best overall for most users” / “If you want the best all-rounder: RoomSketcher.” |
| Copilot | 1.0 | 1.0 | 1.0 | 1.0 | #1 under a “Best Overall Picks” header; cites “ranked #1 by multiple 2026 reviews.” |
| Gemini | 1.0 | 0.75 | 1.0 | 1.0 | Creator: “1. Best Overall & For Real Estate: RoomSketcher.” Software: slips to #2, winning “Best for Real Estate & Interior Design.” |
| Claude | 0.75 | 0.75 | 1.0 | 0.75 | Refuses an overall winner by design (“the ‘best’ lists are almost all vendor-owned content”). Gives RoomSketcher the real-estate / listing-plans category outright in both. |
| Google AI Mode | 1.0 | 0.75 | 1.0 | 1.0 | Creator: “The absolute best floor plan creator software overall is RoomSketcher.” Software: 2nd of three named, leads “Advanced and Professional Software.” |
| Total | 5.75 | 5.25 | 6.00 | 5.75 | 87.5% and 95.8% respectively |
“Software” fell 8.3 points and “creator” 4.2. Three engines — ChatGPT, Perplexity and Copilot — returned a clean 1.0 on both queries, and ChatGPT alone drives 78% of AI referral traffic, so the commercially dominant surface is fully intact. The slippage is two single-step moves: Gemini and Google AI Mode each dropped from “best overall” to “best for real estate / professional” on the software phrasing.
/floor-plan-creator/ page was built for./use/free-floor-plan-creator landing page and a DeskFlex roundup segmenting by use case (offices, classrooms, restaurants, weddings).AI Mode was captured twice per query this cycle, minutes apart, in identical conditions. The answers were not variations on a theme — they were different recommendation sets. On “software” the first run omitted RoomSketcher from all four categories (scoring 0.25) and the second placed it second of three named tools (0.75). On “creator” the first gave Floorplanner the “Best Overall for Beginners” slot (0.75) and the second opened with “the absolute best floor plan creator software overall is RoomSketcher” (1.0).
All four spelling permutations were captured this cycle (24 answers total), which allows a direct test of whether the one-word/two-word distinction is real. It is — but only on some engines.
| Query | Spelling | Score | Status |
|---|---|---|---|
| best floorplan software | one word | 87.5% | Locked anchor — trends from June 1 |
| best floor plan creator | two words | 95.8% | Locked anchor — trends from June 1 |
| best floor plan software | two words | 87.5% | New series — baseline set this issue |
| best floorplan creator | one word | 87.5% | New series — baseline set this issue |
The head terms improved this cycle — in several cases substantially. That makes the traffic decline easier to diagnose and the remaining gaps more valuable: the terms RoomSketcher still has no dedicated page for are the ones where competitors and AI-native entrants are unopposed.
A dedicated /floor-plan-creator/ page is now live — titled “Floor Plan Creator – The Easiest Way to Make Floor Plans,” with a clean H1, a proper meta description referencing AI-powered tools, and a full page build. This closes a gap that had been open across every prior issue.
Verified this cycle: /floor-plan-maker/, /floor-plan-app/, and /ai-floor-plan-generator/ all return 404. RoomSketcher ranks 24.1 on “floor plan maker” on the strength of unrelated pages alone. Meanwhile floorplanmaker.ai now exists as an exact-match AI-native domain on that term. The “creator” page took multiple cycles to ship and is already showing a 5× impression lift — “maker” is the same trade at 19 impressions/day of standing demand, and the competitive window on it is closing.
Six of the eight tracked head terms improved position this cycle, several sharply: “floor plan” 12.5→7.4, “free floor plan software” 10.1→5.5, “room planner” 29.9→20.7, “floor plan app” 10.0→8.9, “floor plan software” 9.6→8.8. Only “floor plan creator” slipped, and that has a page-consolidation explanation. The domain's authority on its core vocabulary is strengthening even as total clicks fall — which is exactly why the traffic decline should not be read as an SEO failure.
The established competitor set is stable. The change this cycle is underneath it: a cohort of AI-native tools has consolidated around “describe your space in plain English→get a floor plan,” and 2026 round-up content now treats text-to-floor-plan as a baseline expectation rather than a differentiator.
| Competitor | “Creator” Page | “Maker” Page | AI Feature | Positioning / Differentiator |
|---|---|---|---|---|
| RoomSketcher | Yes — new | No (404) | Yes | Professional results, beginner-friendly; deep template gallery; real-estate and listing workflows |
| Cedreo | Yes | — | Not AI-branded | Home builders, contractors, remodelers. “Draw a complete house in 2 hours.” Room-specific planners (kitchen, bathroom, basement) |
| Planner 5D | Yes | Yes | Yes | Free tier + consumer-friendly; cited in 2026 AI round-ups as a free AI option |
| Floorplanner | Yes | Yes | Partial | Volume/embed play for real estate portals |
| SmartDraw | Yes | Yes | Partial | Diagramming breadth; strong exact-match SEO footprint |
| SketchUp | — | — | Partial | Professional 3D modelling; architect/designer end of the market |
| floorplanmaker.ai | — | Exact-match domain | Native | New entrant. Free text-to-floor-plan generator on the exact term RoomSketcher has no page for |
| Maket.ai | — | — | Native | Named top residential AI generator in 2026 round-ups |
| Snaptrude | — | — | Native | Text brief → LOD 300–350 BIM, exports to Revit. Professional/architect end |
The new cohort — floorplanmaker.ai, Maket.ai, ArchyBase, Snaptrude, TestFit, Finch3D, RoomGPT — is not currently taking RoomSketcher's rankings; the head terms improved this cycle. The threat is narrower and more specific: these tools are building their SEO and their LLM presence on “AI floor plan generator” and “floor plan maker”, and RoomSketcher has no page on either. When an assistant assembles a recommendation set for “AI floor plan generator,” there is nothing on roomsketcher.com built to be cited for it — despite the product having AI features and the new creator page's own meta description advertising them.
Both Claude responses in this cycle's LLM Visibility capture volunteered unprompted competitive detail worth acting on. Matterport published an “8 Best Floor Plan Software” piece in February and refreshed it days before this reading; Claude characterised its framing as deliberately anti-listicle — arguing there is rarely one best tool, only the best tool for a specific job — which it described as “a smart wedge against RoomSketcher's own 'RoomSketcher is best overall' positioning on the same keyword.” Space Designer 3D and My Site Plan are running comparison plays on the same term.
A spot-check of Cedreo's homepage this cycle found no AI-branded features and no new product launches. Positioning remains squarely on home builders, contractors, remodelers and interior designers — the professional segment — with speed as the core claim. Cedreo is not currently contesting the homeowner/DIY end where RoomSketcher's gallery and free tier compete, and it is not yet using AI as a wedge. That leaves the AI framing available to whoever claims it first.
Four things held or improved this cycle against a broadly negative backdrop. Each is worth protecting explicitly, because in a down cycle the instinct is to change things that are working.
Gallery clicks held at exactly 90.6/day across both windows while site-wide organic clicks fell 16.5%, and the number of gallery pages earning clicks rose from 157 to 166. With 832 pages drawing impressions, this is the most resilient organic asset on the property. It is the single clearest argument for continuing to invest in gallery depth — it is the part of the site that did not decline.
Account-level CTR improved from 8.27% (last reported) to 9.84% — a meaningful gain that indicates ad relevance and query matching are getting better, not worse. Clicks rose 34% on the baseline. Whatever is wrong with paid efficiency this cycle, it is not the ads' ability to earn a click. The problem sits downstream, in what those clicks are optimised toward.
Microsoft's blended cost per conversion fell from $12.82 to $11.50 as spend was cut 33%. Two campaigns improved materially: Branded NORAM CPA $138→$93, and Competitors NORAM $296→$227 — the latter still expensive but moving the right way. The pullback was executed cleanly on the efficiency metric; the cost was volume, particularly in subscriptions.
begin_checkout rose 1.3% and sign_up fell only 3.1% — against a 16% organic session decline. All-source sign_up is still 11% above the last reported period (274→304/day). The top of the funnel narrowed but intent-to-buy did not, which suggests the underlying product demand is intact and this is a traffic-and-efficiency problem rather than a market problem.
The PageSpeed Insights API quota was exhausted again this cycle — the second consecutive issue — so Lighthouse was run locally instead. This produces real, current measurements, but under materially more favourable lab conditions than PSI's throttled environment. Read the numbers below as a fresh absolute read, not as a trend against Issues #1–#7.
Issues #1–#7 reported PageSpeed Insights values (most recently mobile Performance 60, mobile LCP 13.1s, carried forward unverified since June 15). This issue reports local Lighthouse 12 runs on an unthrottled Mac over a residential connection. That environment is much faster than PSI's simulated mid-tier mobile device, so the scores are structurally higher and a comparison between them would be meaningless. No delta is shown for that reason. What these runs can establish is the current absolute state and, more usefully, the variance.
The variance is the finding. Four runs of the same URL, same machine, minutes apart, produced Performance scores of 66, 87, 93 and 81 and LCP values of 10.85s, 3.29s, 2.88s and 4.72s. A stable site does not do this. It means some mobile visitors are getting a genuinely fast homepage and others are waiting three to four times longer, and which one you get appears close to random.
wpmedia.roomsketcher.com — a separate host from the main domain, which adds DNS + TLS negotiation before the largest element can even begin downloading.The PSI API returned a daily-quota error on both the keyed and unkeyed endpoints during this run, exactly as it did on June 29. The consequence is that the mobile-speed regression first flagged in Issue #5 has now gone two full cycles without a comparable measurement, and the 13.1s figure still being carried in prior issues is nine weeks stale. Local Lighthouse is a usable stopgap but it cannot rebuild the trend line. Re-running PSI early in the next cycle, before the daily quota is consumed, is the only way to restore a comparable trend line.
Five items below are new or materially changed this cycle. The paid-efficiency finding is first because it is the largest controllable sum of money on the table, and because it resolves to a specific campaign on specific dates rather than a diffuse "auction got harder" story.
All four NORAM campaigns run Maximize Conversion Value with an explicit target ROAS. Not one target is set at or above the 1.0 the Q3 KR requires:
| Campaign | Target ROAS | Actual ROAS | Daily Budget | Impr. Share | Lost to Budget |
|---|---|---|---|---|---|
| 0. S – Branded | 0.94 | 1.09 | $1,638 | 79.9% | 0.0% |
| 2. S – Competitors | 0.74 | 0.68 | $528 | 18.0% | 5.8% |
| 3. P-Max | 0.74 | 0.68 | $136 | 14.8% | 8.6% |
| 1. S – Search | 0.80 | 0.27 | $228 | 17.4% | 26.6% |
| Combined US (incl. Microsoft) | — | 0.60 | — | 31.8% | — |
Smart Bidding is doing what it was told. Hitting the OKR requires changing the instruction, not the execution — the targets are the KR, expressed in the only place the algorithm reads.
An earlier draft of this issue flagged 1. S - Search - NORAM as money left on the table: it was pinned at ~$1,530/day on July 21–23 at $5.78–$8.13 per conversion — the best CPA in the window — then collapsed to $54–$204/day. Conversion-value data inverts that reading.
Combined Google + Microsoft spend rose 28% ($1,555→$1,984/day) while combined new subscriptions fell 4% (9.12→8.75/day), taking cost per paid subscription from $171 to $227. The two engines failed differently and neither offset the other:
Of 232 new_subscription events in the current window, 205 landed in GA4's “Unassigned” channel — along with 212 of 432 purchase events — on just 1,172 sessions. Direct received 16, Paid Search 6, Organic 4, AI 1. In practice this means GA4 currently cannot answer "which channel drives paid subscriptions," which is the single question that matters most for budget allocation.
user_id but lost session source — most likely a cross-domain or app-handoff gap between the WordPress site and the Flutter app.new_subscription rose from 187 to 205 between windows.Section 2 splits the five-month decline into its parts. The split matters because it changes what to do about it.
These were open at Issue #7. The Q3 OKRs pick up two of them directly, which is worth noting explicitly — they move from "gap" to "scheduled."
Two external shifts frame everything above: search is resolving more queries without a click, and text-to-floor-plan generation has become a table-stakes feature rather than a novelty.
2026 comparison content now treats "describe your space in plain English and get a to-scale plan" as an expected capability, alongside sketch-to-plan and photo-to-plan. Tools are differentiating on what happens after generation — Snaptrude on BIM/Revit export, TestFit on multi-family feasibility, ArchyBase on connecting plans to rendering and 3D. RoomSketcher's genuine differentiators (template depth, real-estate listing workflows, professional output from a beginner-friendly editor) are durable, but they are increasingly hard to discover through AI answers if the site has no page that speaks the AI vocabulary.
Google's July 9 confirmation that smaller core updates now roll without announcement removes the old playbook of "wait for rollout to finish, then assess." There is no rollout end-date and no named update to recover from. The practical adaptation is to monitor CTR-at-stable-position as its own metric — which is precisely the pattern visible in this cycle's data — and to treat answer-surface presence as the hedge against it.
Six items for the next 14 days, ordered so the diagnostics and confirmations that unblock everything else come first. Items marked KWORQ are ours to execute. Objective tags map each action to the Q3 OKR it serves.
All four NORAM campaigns run Maximize Conversion Value with an explicit target ROAS set below the Q3 KR: Branded 0.94, Search 0.80, Competitors 0.74, P-Max 0.74. The account is delivering 0.60 because that is roughly what it was asked for. Our recommendation is to raise the targets toward 1.0 in staged increments rather than in one step — Smart Bidding needs a learning window and volume will dip before efficiency recovers — starting with Competitors and P-Max at 0.74. Branded is already returning 1.09 and is best left alone. Happy to model the likely volume trade-off at each target level if that would help the decision.
1. S - Search - NORAM went from ~$1,530/day to a configured $228/day around July 24–27 — an ~85% cut. On ROAS that was the right move (the campaign returns 0.27, the worst in the account). We need to know whether it was an intentional response to poor return or an unrelated budget action, because it determines whether the same value-based scrutiny is already being applied to Competitors and P-Max, both sitting at 0.68.
88% of new_subscription events land in "Unassigned." Trace where session source is lost between the WordPress site and the Flutter app — most likely a cross-domain linker or app-handoff gap — and restore channel attribution on revenue events. Note the scope: per-engine ROAS is measurable from platform data (Section 0), so this does not block the Objective 1 KR. What it blocks is comparing paid against organic and AI on the same basis — which is what the budget-allocation decision actually needs.
Objective 3's KR needs a measurement method robust enough to score against. Three changes, all evidenced this cycle: capture Google AI Mode three times per query and score the median (two runs minutes apart produced different recommendation sets); track all four spelling variants rather than two, since Copilot ranks the one-word and two-word forms differently and the two-word form carries 12.7× the search volume; and verify the query string in every screenshot before scoring. Baselines for all four are set in Section 3.
The creator page shipped and produced a 5× impression lift within two weeks. "Floor plan maker" carries 19 impressions/day of standing demand at position 24 with no dedicated page behind it, and floorplanmaker.ai now holds the exact-match domain. Same trade, same evidence, one term over.
The product has AI features — the new creator page's own meta description says so — but /ai-floor-plan-generator/ is a 404 and nothing on the site is built to be cited when an assistant assembles recommendations for that phrasing. This is the term the entire new competitive cohort is organising around. The decision needed this cycle is whether to claim it; the build itself can follow.
| Term | Meaning |
|---|---|
| SEO | Search Engine Optimisation — earning visibility in conventional search results. |
| AEO | Answer Engine Optimisation — being the source an AI assistant cites when it answers directly. |
| GEO | Generative Engine Optimisation — shaping how generative AI systems represent and recommend a brand. |
| GSC | Google Search Console — source for impressions, clicks, CTR and average position. |
| LCP | Largest Contentful Paint — time until the largest visible element renders. Google's "good" threshold is 2.5s. |
| CLS | Cumulative Layout Shift — how much the page moves while loading. Below 0.1 is "good". |
| TBT | Total Blocking Time — how long the page is unresponsive to input during load. |
| CTR | Click-Through Rate — clicks divided by impressions. |
| CPC | Cost per Click. |
| CPA | Cost per Acquisition — spend divided by conversions. In this report, "blended" CPA counts all conversion actions including free accounts. |
| P-Max | Performance Max — Google's automated campaign type spanning Search, Display, YouTube and Shopping inventory. |
| Unassigned | GA4's bucket for sessions where the traffic source could not be determined — see Section 8. |