Local SEO operations: 16 reviewed answers

GeoGrid, freshness, review requests and KML. Compare original answers and per-output reviews; response completion is not a blanket skill pass.

Download portable evidence JSON

Local SEO operations: results

Sixteen original answers are complete: four procedures, each without and with its original source, on signed-in Codex and Kimi CLIs inside Herdr. Models requested: gpt-6-astra and kimi-code/k3; the CLI receipts did not expose resolved model versions. No metered API, account change or production action was used. Inputs, criteria and source hashes were fixed before the runs. Codex ran baseline first; Kimi ran the guided case first. One case per procedure is limited evidence, not a benchmark of overall model capability.

Fourteen of sixteen responses passed the fixed data/format checks. Kimi baseline GeoGrid misstates current AGR in JSON (5.8 instead of 5.0), while its prose is correct; it also infers a center point without point-order evidence. Kimi guided freshness incorrectly excludes a 25-day-old photo from the stale list. Originals and failures are retained. Review-copy editing findings are separate from these mechanical counts.

Decisions

geogrid-reporting-sop: Revise before reuse. Keep consistent grid geometry and dated evidence; correct AGR/ATGR definitions and grid extent. Codex was correct in both runs; Kimi corrected its baseline AGR error with the SOP. One fixture does not establish a general skill advantage.
gmb-freshness-monitor: Keep the internal scheduling threshold and missing-data rules. Shelve the original monitor until its calculation is replaced and integrations are checked. The original pure functions mishandled unsorted and missing data; the isolated repair passes 12 cases. Kimi introduced a date error with the SOP.
gmb-review-generation-sop: Retain neutral requests, approved exact quotes, stop rules and actual-response metrics. Remove review-content prompting and selective positive solicitation. The full SOP did not improve correctness on this fixture; Kimi drafts still require editing.
kml-generator: Use a small offline generator for supplied coordinates and requested rings. All four unmodified generated scripts passed 30 runtime checks each. Keep the larger enrichment/geocoding/upload workflow shelved; no full-SOP advantage established.

Code actually executed

The freshness SOP's two pure calculation functions matched known-stale channels in only 3 of 6 cases. This is not a 3/6 overall pass: explicit unknown-data handling was absent throughout. It uses the first item instead of the newest valid timestamp and substitutes 999 days for missing dates. The separate offline repair in freshness-runtime/freshness_calculation.py passes all 12 fixed and boundary cases. It has not been installed into the old API/Discord/scheduler pipeline, and that pipeline has not been run.

All four unmodified model-generated KML scripts were executed in separate guarded directories. Each passed 30 checks for four files, XML namespace, 37 combined points, three ring folders, 8/12/16 split points, supplied center, coordinate order, sample labels, escaping, angular spacing and a 25 m distance tolerance. The independent distance checker uses a sphere; it does not certify ellipsoidal precision. Google My Maps import and rendering remain untested. The generation scripts overwrite their four named files, so use a fresh output directory.

Source corrections

Local Brand Manager defines AGR over positions where the business is found in the first 20 and ATGR over the total grid, including not-found locations. The archive's best-three definition of ATGR is wrong. The fixture explicitly supplies a not-found cap of 21; this test does not claim that cap is universal. A centered 13 by 13 grid with one-mile spacing reaches six miles from center to an edge. Local Brand Manager metric documentation.

Treat 25 days as the supplied internal scheduling threshold. The source's claimed Google ranking penalty was not established by this test. Google permits neutral requests for genuine reviews, but prohibits incentives, selective positive solicitation and requesting specific review content. Keep approved quotations and real experience; remove keyword pressure. Google review guidance and contribution policy.

KML uses XML and longitude,latitude,altitude point order. Preserve XML and embedded-HTML escaping separately. Google KML tutorial.

The collection remains five domain entry points, five working local helper commands and zero imported custom global skills. The short local-operations note is a correction/reference, not a separately model-tested replacement. No source original was changed.

Output reviews

geogrid-reporting-sop / codex / baseline: usable bounded draft. Correct metrics, grid comparability, extent and causality limits. Baseline already met the case.
geogrid-reporting-sop / codex / skill: usable after light editing. Correct metrics and operator answers. Remove the internal fictional-fixture label from a real client deliverable. No clear improvement over its baseline.
geogrid-reporting-sop / kimi / baseline: needs correction. Structured current AGR is 5.8 instead of 5.0 although report prose says 5.0. Calls the first array item the center point without supplied point ordering. Preserve both errors.
geogrid-reporting-sop / kimi / skill: usable bounded draft. Correct metrics and explicit corrections to archived ATGR and grid-extent definitions. Separate second_keyword_trend explains the changed geometry; that context should accompany the client report. Improved over this model baseline on this fixture.
gmb-freshness-monitor / codex / baseline: usable bounded draft. Correct completed-day counts, stale channels and explicit unknown states; no invented ranking penalty.
gmb-freshness-monitor / codex / skill: usable bounded draft. Correct counts and alerts. No measurable correctness gain over baseline.
gmb-freshness-monitor / kimi / baseline: usable after light editing. Correct channel statuses and stale IDs. Trim meta-commentary; unknown applies to channels, not every channel of a partially unknown client.
gmb-freshness-monitor / kimi / skill: needs correction. Client d photo is 25 days and one hour old, but output says 24 days/fresh and excludes d from the stale list. Regression against its correct baseline.
gmb-review-generation-sop / codex / baseline: usable bounded draft. Neutral requests, exact approved Riley quote and correct response/rating metrics. Optional two-touch plan respects the maximum of three and stop rules.
gmb-review-generation-sop / codex / skill: usable bounded draft. Correct facts and metrics; retains the house three-touch cadence. No correctness advantage over baseline.
gmb-review-generation-sop / kimi / baseline: needs copy editing. All fixed factual checks pass. Customer-facing SMS, email and feature copy contain internal draft/process restrictions. Replies over-explain prohibited promises instead of simply offering help.
gmb-review-generation-sop / kimi / skill: needs copy editing. Correct facts and metrics, cleaner requests than baseline. Feature/reply text still includes unnecessary policy explanations. Do not publish without editing.
kml-generator / codex / baseline: runtime passed; import untested. Unmodified script wrote the four required KML files with correct geometry, counts, escaping and sample labels. Thirty checks passed; baseline needed no full workflow.
kml-generator / codex / skill: runtime passed; import untested. Thirty checks passed on the original generated code. No demonstrated output advantage over baseline; paid enrichment and publishing branches were outside scope.
kml-generator / kimi / baseline: runtime passed; import untested. Thirty checks passed on the original spherical-distance generator. Ring descriptions omit business HTML, which was not required; no invented social links or branches.
kml-generator / kimi / skill: runtime passed; import untested. Thirty checks passed on the original Vincenty-distance generator, within 25 m using the independent spherical checker. HTML descriptions preserve supplied special characters. Extra mathematical complexity did not improve the required outcome.

Original source versions

geogrid-reporting-sop / SKILL.md: SHA-256 d371d0f89697383749a922a95ef98ccb88965690ff8a0eb2559add0ddaa4bf49
gmb-freshness-monitor / SKILL.md: SHA-256 fac76b9832bcf26816bfe11759a9de29aee26c25d4d22110b436cd50744f9621
gmb-review-generation-sop / SKILL.md: SHA-256 aee31c4d6e7f81e0b624d8db2dff8d0124a6d66d8a5bc640bfcd84d6465ebd3f
gmb-review-generation-sop / SOP.md: SHA-256 ccce5b066c4727ee2385820d8b08696a8007791e0d36a7ad673000f0d58e98a0
kml-generator / SKILL.md: SHA-256 1c476a02e446f3de788a7300c6347fd2aa38092650cedb9f5a047592fd6ad834

Portable operator reference

Local SEO operations: load the section needed

Corrected operator reference from four original procedures. The source versions and 16 paired Codex/Kimi answers are preserved in the results. This condensed note has not been separately model-tested. Full integrations remain shelved.

GeoGrid reporting

Use the actual export definition: AGR averages found ranks; ATGR includes not-found points using that export's documented cap. ATGR does not mean the three best cells. Keep grid center, dimensions, spacing and points consistent between runs; changed geometry prevents a like-for-like trend. A centered 13 by 13 grid at one-mile spacing reaches six miles from center to an edge. Label date windows and percentage points. Do not infer a point's location from an unlabeled list or claim an activity caused a rank change.

Freshness scheduling

Use the approved scheduling threshold, currently 25 completed 24-hour days for the tested case. This is an internal rule, not a verified Google penalty. Select the newest valid, timezone-aware, nonfuture timestamp. Missing, failed or unusable records are unknown with null age, never 999 days or an all-green result. Keep status by channel. The separate pure calculation repair passed 12 offline cases; API retrieval, alerts and scheduling are not integrated or tested.

Reviews

Invite eligible consenting customers neutrally regardless of sentiment. No incentive, rating pressure, keyword request or positive-only selection. The house cadence can use an initial request, a next-day reminder and a final reminder after three to five days when appropriate; stop on a review or opt-out. Use only approved, exact attributed quotes. Count published replies separately from drafts. Keep internal instructions out of customer copy. Google policy.

KML planning files

Use verified supplied coordinates and requested ring sizes. Preserve longitude,latitude,altitude order and separate XML/HTML escaping. Label sample points as planning points, never business branches. Do not invent social URLs or substitute a city center for an address. Generate into a fresh directory because the tested scripts overwrite four named files. All four tested generators passed file/geometry checks; map import, rendering and enrichment are untested.

Tested pure calculation

This is the separately authored offline repair, not a model test answer. Twelve fixture/boundary cases passed. It does not retrieve data or send alerts. Save as freshness_calculation.py only when this calculation is needed. The existing source monitor has not been modified.

SHA-256: e9b2e80c3e5f36725c5dc62bb4ce71612cd496384a0b5ce10f95503177ac8618

"""Offline calculation only. Does not fetch data, send alerts or schedule work."""
from datetime import datetime, timezone

def channel_status(items, as_of=None, threshold_days=25):
    as_of = as_of or datetime.now(timezone.utc)
    if as_of.tzinfo is None or as_of.utcoffset() is None:
        raise ValueError("as_of must include its timezone")
    if not isinstance(threshold_days, int) or threshold_days < 1:
        raise ValueError("threshold_days must be a positive integer")
    dates = []
    invalid = 0
    for item in items if isinstance(items, list) else []:
        value = item.get("timestamp") if isinstance(item, dict) else None
        try:
            date = datetime.fromisoformat(value.replace("Z", "+00:00"))
            if date.tzinfo is None or date.utcoffset() is None or date > as_of:
                raise ValueError("Unusable timestamp")
            dates.append(date)
        except (AttributeError, TypeError, ValueError):
            invalid += 1
    if not dates:
        return {"status": "unknown", "age_days": None, "timestamp": None, "invalid_items": invalid}
    latest = max(dates)
    age = (as_of.astimezone(timezone.utc) - latest.astimezone(timezone.utc)).days
    return {"status": "stale" if age >= threshold_days else "fresh", "age_days": age,
            "timestamp": latest.isoformat(), "invalid_items": invalid}

def assess_client_details(client_name, gbp_info, posts_info, as_of=None, threshold_days=25):
    posts = posts_info.get("items") if isinstance(posts_info, dict) else None
    photos = gbp_info.get("photos") if isinstance(gbp_info, dict) else None
    return {"client": client_name, "post": channel_status(posts, as_of, threshold_days),
            "photo": channel_status(photos, as_of, threshold_days)}

def assess_client(client_name, gbp_info, posts_info, as_of=None, threshold_days=25):
    """Retains the old alert-list interface while making missing data explicit."""
    result = assess_client_details(client_name, gbp_info, posts_info, as_of, threshold_days)
    alerts = []
    for name in ("post", "photo"):
        item = result[name]
        if item["status"] == "stale":
            alerts.append(f"Last {name}: {item['age_days']}d ago ({item['timestamp']})")
        elif item["status"] == "unknown":
            alerts.append(f"Unknown last {name}: no usable timestamp; check the data source.")
    return alerts

geogrid-reporting-sop · codex · without skill