The LSA Answer Rate Everyone Asserts and Nobody Publishes

Every 2026 guide to Google Local Services Ads names response speed as a ranking signal. None of them shows the measurement — so here is one.

Below is a per-hour answer rate for a live Local Services channel, measured against the same account's weekly reach and cost-per-lead series over the same seven weeks. The finding is that this account's reach collapsed by 94% in four weeks while its price per charged lead held flat within three dollars, that no configuration change occurred inside the window where the collapse happened, and that the one input the account does not control directly — how many of its inbound calls get answered — was at its worst in the week the decay began.

That is evidence, and it is not proof. Answer rate driving ad rank is Google's documented model and every guide in the category repeats it, but the causal link to this collapse is unproven and this post does not report it as established. No API field exposes LSA ad rank; it is console-only, which means the intermediate variable in the chain cannot be read at all. What we can do is publish the measurement, publish the reasoning, and publish the bar that would falsify it. The account is ProComfortSolutions, an HVAC contractor in Chicago and an NBM client.

Lead with the hours, not the average

Daypart	Calls	Answered ≥30s	Answer rate
08:00 – 12:00	11	7	64%
12:00 – 17:00	19	6	32%
17:00 – 22:00	15	8	53%

Population and method: 46 inbound calls to the Local Services channel since 2026-06-29, every one filtered to consumer participants — advertiser callbacks share the same channel and were excluded before any rate was computed. Measured 2026-08-11. "Answered" means a connected call of at least 30 seconds. The three dayparts above account for 45 of the 46 calls.

Blended, that is 21 of 46: 46%. The blended figure is the least useful number in the table. It hides a two-fold spread, and it points at the wrong shift — the busiest block of the day answers at half the rate of the quietest one. Of the 25 calls that were not answered, 24 rang for 1–29 seconds and one had zero duration. 24 of those 25 have a Google call recording attached, which makes them auditable one at a time rather than statistically.

What happened to reach

Week	Impressions	Clicks	Cost	Charged leads
W27	2,106	110	$1,109.71	13
W28	1,931	93	$546.70	6
W29	4,767	199	$1,564.95	19
W30	2,841	110	$560.75	7
W31	1,321	65	$555.01	7
W32	1,185	47	$110.22	1
W33	294	9	$0.00	0

Read 2026-08-11 from the Local Services lead resource and campaign reporting. 94% of reach lost in four weeks from the W29 peak. Daily budget over the whole series: $214.29, and unspent. Configuration intact at the time of the read: 7 ad schedules, 48 locations, 12 service categories.

Separating the Price Signal from the Rank Signal

Step 1 — Plot reach and cost per lead on the same weekly axis

Most reach diagnoses stop at the impressions line and reach for a budget explanation. Put price beside it and the diagnosis usually resolves in one glance.

Here: impressions fall 4,767 → 294 across four weeks, while cost per charged lead moves $85.79 → $88.12. Roughly $86 a lead is the benchmark any relaunch on this account gets measured against, because it is the one number that did not move.

Step 2 — If price holds while volume collapses, the auction did not reprice you; it derated you

A budget or bid problem announces itself as a price problem. You get outbid, the clearing price rises, your volume falls, and cost per lead rises with it. That is not what this account did. The price is flat within three dollars across a collapse that removed nineteen twentieths of its reach, and the budget was never exhausted.

The account was not outbid. It was shown less. Those are different failures with different remedies, and raising the budget — the reflex response to a volume drop — would have spent more money to buy the same nothing.

Step 3 — Rule out every configuration cause via the change log

The change-event log for the Local Services campaign over 2026-07-19 → 08-18 returns exactly two touchpoints: 2026-07-20 07:18:30 and 2026-08-11 19:02:11. Nothing changed between 2026-07-21 and 2026-08-10, and the cliff sits inside that untouched window.

One caveat, and we have to raise it because we raised it ourselves in a different post about the same account: a silent change log is weak evidence of absence. That surface has scope boundaries — conversion-action edits, for instance, do not appear in it at all. What the two touchpoints do establish is narrower and still useful: no campaign-level configuration edit occurred inside the window in which reach collapsed.

The 08-11 edit is worth naming precisely, because it looks like a suspect and is not one. It cut weekday evenings from 22:00 to 20:00, leaving the schedule at Mon–Fri 07:00–20:00. It post-dates the cliff by 13 days. It removed two evening hours from a channel selling emergency HVAC, which is a real cost — and it is not the cause, and restoring it would not address the deficit the measurement actually found.

Step 4 — Look for the ranking input you do not control directly

Budget, bid, schedule, service categories and geography are all inputs an advertiser sets. Every one of them was intact and unedited across the collapse. That exhausts the levers on the account side and moves the search to inputs Google states it observes and the advertiser does not set.

Responsiveness is the one in that category with a published model and a measurable proxy — and the proxy is sitting in the call log, timestamped, with recordings attached.

The daily series, and the day it stopped

Live-read 2026-08-17 from the lead resource: 2026-07-01 → 07-28 covered 22 active days and 67 leads, peaking at 7 on 07-02, typically 2–5 a day. Then: 07-29 zero. 07-30 one. 07-31 → 08-04 zero. 08-05 one. 08-06 zero. 08-07 one. 08-08 → 08-13 zero. 08-14 one. 08-15 → 08-17 zero.

The last normal day is 2026-07-28. Three leads in the nineteen days since.

The week that looks like the trigger

W29 is the peak of the reach series at 4,767 impressions. It is also the worst answer week in the set: 13 consumer calls, 4 answered, measured 2026-08-11 on the same consumer-filtered population as the daypart table. That is the call population, not the charged-lead column above; the two are different populations and are not compared here. The highest-volume week carried the lowest answer rate, and reach decayed every week after it.

That is a correlation with a plausible mechanism, a documented vendor model behind it, and an untouched configuration window around it. It is not a demonstrated cause. Volume and answer rate are not independent — a spike in calls is exactly the condition under which a small office starts missing them — so the direction of the arrow between "busy week" and "bad answer week" is not settled by this data either.

What the guides say, and what that is worth

Blue Grid Media's 2026 LSA ranking guide is direct about it: "If you miss calls or take hours to respond to messages, the algorithm records this and adjusts your placement downward." It adds that "Businesses with a high phone response score consistently appear in the top 3 LSA positions" and, usefully for anyone who thinks a fast callback repairs the damage, that "Calls that go to voicemail count as 'not answered' in Google's system, even if you call back within minutes." SmartSites' 2026 overview of Local Services Ads covers the same ground for the same audience.

These are assertions, and they are probably right. Neither publishes a per-hour answer rate against a reach series from the same account over the same weeks, which is the artifact the category is missing. Filling that gap is what this post does. It does not close the causal question, and treating a published assertion as a substitute for a measurement is the habit that produced years of confident LSA advice with nothing measured underneath it.

The premise we published, then refuted with our own measurement

The plan written for this account before the measurement existed asked whether overnight calls were going unanswered. That was the wrong question and the measurement says so plainly. The evening block answers at 53% and the morning block at 64%. The hole is midday — 12:00 to 17:00, at 32%, on the highest call volume of the day.

Which is why the recommendation here is not to extend Local Services hours. Extending hours adds calls to a business that is already missing two-thirds of the calls arriving in the block it is fully staffed for. The measured deficit is a staffing-coverage problem inside existing hours, and buying more hours on top of it would make the number worse while looking like action.

Three ways this channel's instruments lie

Anyone attempting to reproduce this measurement will hit all three.

1. leadCharged is omitted when false. Proto3 omits default values, so counting rows where the key is present returns 68 July leads instead of 49 — an overstatement of 39%. Test for the value, never for the key.
2. segments.date is a prohibited segment on the lead resource. Filter on the creation timestamp instead. The query fails loudly, which is the merciful version.
3. The campaign metrics view disagrees with the lead resource. metrics.conversions read 0 for 2026-08-14 while the lead resource returned a row for the same day. For Local Services lead truth, the lead resource is authoritative and the campaign metrics view is not reliable for this channel.

Three charged calls that nobody answered

Beyond the ranking question there is a smaller, entirely concrete one. Three charged Local Services leads are verified as calls nobody answered, each with a Google call recording attached. They are disputable individually, on evidence, rather than as a blanket complaint.

If a figure is wanted: approximately $255 across three verified charges. An earlier internal estimate was higher and has been corrected downward, because it rested partly on a lead that had already been credited. A fourth candidate remains unchecked. The smaller, verified number is the one that goes in public.

The June outage was something else entirely

A separate blackout ran 2026-06-17 → 06-29 and has a fully documented, unrelated cause: the account's insurance artifact failed verification three times — EXPIRED against a document expiry of 2025-03-17, then NO_SIGNATURE on 2026-06-15, then OTHER on 2026-06-26 — and finally passed on 2026-06-30, the day after the blackout ended.

That is a different population with a different mechanism, and folding it into the July cliff would mix the two. It is in this post only so that nobody re-reading the reach series mistakes June for part of the same story.

What would prove this wrong

This is the section the rest of the post is written to earn.

• The bar we have committed to: answer rate at 80% or better, sustained 14 consecutive days, then a re-read of weekly impressions. If the answer rate clears 80% and holds, and reach does not recover, responsiveness is not the lever on this account and we will publish that result with the same population, formula and window as this one.
• A Google-side policy, category or eligibility change inside 2026-07-21 → 08-10. If one surfaces, the untouched-window argument collapses and the configuration explanation returns.
• A comparable account, in the same trade and metro, with a comparably poor answer rate that held its reach across the same weeks. That would make this correlation coincidental.
• LSA ad rank becoming API-readable and reading flat across the collapse. Then the derating hypothesis is dead and the mechanism is something upstream of rank entirely.
• A demonstrated reverse direction — reach falling first and call volume falling with it, dragging the answer rate down as a symptom rather than a cause. The W29 sequencing argues against it; it does not rule it out.

Where this leaves the category

The honest summary is short. Reach collapsed, price did not, configuration did not change, and the one uncontrolled input we can measure was at its worst the week the decay started. That is the strongest published evidence we are aware of on a question the whole category asserts and nobody has instrumented — and it is still one account, one channel, seven weeks. If you run Local Services Ads and have never split your answer rate by daypart, the split is one query against your own call log, and it will name the shift that is costing you. We will publish the 14-day read either way.