Invisible to #1
— in 3 days.
Measured on the prompt “Recommend a good digital advertising agency in Phoenix” — audit dated 2026-07-16, three days after the machine layer went live at 2026-07-13 08:45 UTC. Perplexity #1, Grok #1, ChatGPT #2.
On 17 June 2026, five AI engines were asked which digital marketing agencies to consider for enterprise SEO. None of them named V Digital Services. The AnswerShare machine layer went live at the edge of www.vdigitalservices.com on 13 July.
Three days later, the audit above returned V Digital Services first on Perplexity, first on Grok, and second on ChatGPT. On 2 August, Perplexity returned it first again. On 22 July, ChatGPT did not return it at all.
Every statement on this page carries its date and its source, the 22 July dip included. The full before-state, the exact queries behind each measurement, and the figures we withheld are all below.
Go-live is 13 July 2026, and the record has a wrinkle worth naming.
Three timestamps describe this deployment and they do not all mean the same thing. Content preparation, the announcement, and the first measured crawler request are separate events. We anchor go-live to the last of them, because it is the only one that records a machine actually reading the layer.
- 2026-06-28
Clean-room page ingestion begins
717 pages ingested into the clean-room store; content continued to be touched through 2026-08-06. This is content preparation — it is not cutover, and we do not date the deployment from it.
Clean-room page store, earliest row - 2026-07-10
Edge layer configuration circulated
The setup for the AnswerShare edge layer was sent to the client, described in that email as “timed to our go-live this morning.” No crawler traffic is recorded on this date. We read this as the announcement, not the cutover.
Email, Robert Maynard, 2026-07-10 14:54 UTC - 2026-07-13
First AI-crawler request recorded at the edge — go-live
08:45 UTC. The first of more than 16,800 rows in the crawl log for this client. This is the earliest evidence that an inference crawler read the deployed layer, and it is the date every elapsed-time statement on this page counts from.
Crawl log, client_slug vdigitalservices, first row
Why this matters to the numbers
Dating go-live to 10 July would make the first measured audit six days out. Dating it to 13 July makes it three. We use 13 July — the harder, machine-recorded anchor — and note that it is the less flattering of the two available readings only in the sense that it is the one we can prove.
Absent from nine of ten answers, at the exact moment a prospect is choosing.
Three weeks before cutover, five engines were asked two buyer-intent questions about choosing a digital marketing agency. V Digital Services was omitted by all five engines on the enterprise-SEO query, and by four of five on the Phoenix query — Perplexity named it there. The probe wrote 30 receipt files — 10 screenshots, 10 raw API responses, 10 text transcripts — each raw response carrying a SHA-256 of the engine’s verbatim answer text.
That is the whole of the claim. It is a starting condition, not the first half of a comparison — see the note below the tables.
“Which digital marketing agencies should I consider for enterprise SEO?”
| Engine | Model | Result | Competitors named | SHA-256 of answer text |
|---|---|---|---|---|
| Perplexity | sonar-pro | Omitted | Digital Authority Partners; AMP Agency; Clay Agency | 78c8e9e63e…f4f113 |
| ChatGPT | gpt-4.1 | Omitted | — | b39b18550e…1e3a52 |
| Gemini | gemini-2.5-pro | OmittedReview | — | 7b22e100f4…474ef2 |
| Claude | claude-sonnet-4-5-20250929 | OmittedReview | — | 9f5727a19d…5ca071 |
| Grok | grok-4 | OmittedReview | — | d5d0e91a84…9aa430 |
“What is the best digital marketing agency in Phoenix AZ for SEO and paid search?”
| Engine | Model | Result | Competitors named | SHA-256 of answer text |
|---|---|---|---|---|
| Perplexity | sonar-pro | Named | — | 7e7a0a1625…73c8b0 |
| ChatGPT | gpt-4.1 | Omitted | — | 307411bf1e…004500 |
| Gemini | gemini-2.5-pro | Omitted | — | d2ca88973a…af8541 |
| Claude | claude-sonnet-4-5-20250929 | OmittedReview | — | 4b2938a9cf…549f0f |
| Grok | grok-4 | Omitted | — | ef39d78daf…141bd5 |
The result column is read from the receipts, not from the stored verdicts. The classifier that originally labelled each answer matched only the literal strings vdigitalservices and vdigitalservices.com, and could not match the brand written with spaces. It therefore scored Perplexity’s Phoenix answer as an omission when the receipt names “V Digital Services” in its recommendation table. That is why the Phoenix query reads four of five rather than five of five.
What the hashes are hashes of. Each value above is the SHA-256 of the engine’s verbatim answer text as returned by its API — not of the linked .txt receipt file, which carries a four-line prepended header that is not part of the hashed bytes. A separate file-level checksum is recorded for each .raw.json, computed over LF-normalised bytes.
Four of the ten rows are low-confidence. The receipt file’s own caveat reads: rows marked _review_ are low-confidence heuristic classifications — an empty or refused answer, or a partial brand-token match — and should be confirmed by hand against the receipt before use in a client deliverable. That hand-confirmation has not been done. Three of the five rows on Query 1 are flagged; one of five on Query 2.
Only one row’s competitor names are published, because only one row’s survived hand-checking. The stored competitor lists are heuristic token extractions and are unreliable in both directions. The extraction recorded “Right Agency” and “Match Agency” against ChatGPT on Query 1; neither is a company. Both were lifted from section headings inside the answer text — “How to Choose the Right Agency…” and “Match Agency Strength to Your Needs.” The summary file that names them is wrong, and this page does not repeat it. Rows showing “—” mean not hand-confirmed; several of those receipts do name agencies in their prose. Perplexity’s three are published because they appear in the receipt text.
The files were committed in bulk, twelve days after capture. All 30 landed in a single commit on 2026-06-29 for a capture dated 2026-06-17. There is no incremental commit history to corroborate the timeline, so the case for authenticity rests on the internal self-consistency described above — hashes, provider-specific response shapes, token counts, request-timing spread and screenshot dimensions — rather than on version control.
Capture window. Query 2 ran 2026-06-17 17:40:00 UTC to 17:40:02 UTC; Query 1 ran 17:41:52 UTC to 17:41:56 UTC. Live web search was enabled for all probes. 30 files: 10 screenshots, 10 raw JSON, 10 text.
There is a documented before. There is not yet a matching after.
The baseline ran two exact query strings across all five engines: “Which digital marketing agencies should I consider for enterprise SEO?” and “What is the best digital marketing agency in Phoenix AZ for SEO and paid search?”. The post-cutover audit on 2026-07-16 used a third, different query — “Recommend a good digital advertising agency in Phoenix.”
A change in the answer to a different question is not a measured lift, so this page does not present one. A like-for-like re-run of the two baseline queries has not been performed. Until it is, the honest reading is: a documented starting condition, and a separate documented measurement afterwards on its own question.
What we observed, on the day we observed it.
Rankings in generative engines move. A sequence that only moved upward would be a sequence that had been edited. This one includes the day ChatGPT dropped V Digital Services entirely.
| Date | Observation | Source |
|---|---|---|
| 2026-07-15 | AnswerShare GEO composite reported as moving from 77 at cutover to 89. This is AnswerShare’s own composite metric, not an engine ranking. No engine placed anything on this date. | Email, Robert Maynard, 2026-07-15 12:19 UTC |
| 2026-07-16 | Engine audit on the query “Recommend a good digital advertising agency in Phoenix”: Perplexity #1, Grok #1, ChatGPT #2. Gemini and Claude were not measured on this query. Three days after first recorded crawl. | Client dashboard engine-audit panel, audit dated 2026-07-16 |
| 2026-07-19 | Perplexity, on the query “recommend a digital agency in Phoenix”, run in an incognito browser: V Digital Services cited. The observation was ours, not the client’s — “The citation is important. But look how its opinion of VDS has changed.” | Email, Robert Maynard, 2026-07-19 16:10 UTC |
| 2026-07-22 | ChatGPT did not surface V Digital Services on two prompts that day. On a “best digital advertising agency in Phoenix” prompt: “You didn’t come up.” Later the same day, on a separate run: “It didn’t list you in the top5.” | Emails, Robert Maynard, 2026-07-22 12:29 and 21:17 |
| 2026-07-28 | An internal email carrying the subject line “VDS from 82 to 90”. The figures appear only in the subject line — the message body does not restate or evidence them. Recorded here for completeness, not offered as a measurement. | Email, Kevin Myers, 2026-07-28 00:25 UTC — subject line only |
| 2026-08-02 | Perplexity: V Digital Services #1. The clearest dated first-place result in the record. | Email, Robert Maynard, 2026-08-02 20:53 UTC |
The 2026-07-16 audit, cell by cell
| Engine | Position |
|---|---|
| Perplexity | #1 |
| Grok | #1 |
| ChatGPT | #2 |
| Gemini | —not measured on this query |
| Claude | —not measured on this query |
The measurements, including the ones that are not where we want them.
Metric codes are reproduced as they appear on the client dashboard. Where the dashboard and the underlying store disagree, we say so and withhold the figure rather than choose between them.
Crawler traffic
Trailing 30 days as of 2026-08-08. This is a rolling window — it is not anchored to any capture date elsewhere on this page.
Grand total before cleaning: 16,820. Source: crawl log, client_slug vdigitalservices, queried 2026-08-08.
The GEO composite, as a series
AnswerShare’s own composite score — not an engine ranking, and not a third-party measurement. We present it as a dated series rather than a headline number because the stored values disagree with one another.
| Date | Value | Where it is stored | Note |
|---|---|---|---|
| 2026-07-13 | 77 | Reported as the value “at cutover” | Email, Robert Maynard, 2026-07-15 |
| 2026-07-15 | 89 | Reported same morning | Email, Robert Maynard, 2026-07-15 |
| 2026-07-27 | 90 | quant_metrics + comparison_scores | Both tables agree on 90; they disagree on the underlying Grok and Gemini values (97 vs 100) |
| 2026-07-27 | 82 | quant_metrics, later row, same day | Unreconciled against the 90 recorded hours earlier |
| 2026-08-02 | 92 | quant_metrics, row id 128 | Most recent stored value. The client dashboard does not surface it — it still shows 90. |
The five-model median underlying the 90 was captured 2026-07-27; that is the correct date for that score. A further update ran 2026-08-07, but 2026-07-27 remains the authoritative capture for the composite the dashboard displays.
In band
- INFRAInfrastructure
- 100
- SENSentiment
- +85
- ASQAnswer surface quality
- 78
- RRRetrieval reliability
- 0.991
- RTCRetrieval time cost
- 0.008
- TTFBTime to first byte
- 16 ms
- TTLBTime to last byte
- 17 ms
- SCHEMASchema depth
- 2
Out of band, or withheld
- SGRtarget ≥ 0.25
- 0.211
- LMRtarget ≤ 30 days
- 53 days
- SEOsoft — below the range we consider settled
- 57
- EGRsoft — below the range we consider settled
- 33%
- CITEwithheld
- The dashboard reads 55%; the underlying store reads 60. Not published until the two are reconciled.
- RPCdropped
- The dashboard displays 6,176. No corresponding field exists in the measurement store, so we cannot say what it counts. Removed rather than reprinted.
“Working with the AnswerShare team has fundamentally changed how we think about AI visibility. At V Digital Services, we’ve always focused on helping clients win in search, and we have been very successful at doing it. But it’s a new world. With AnswerShare as a partner, we are now prepared to implement for clients for where “search” is heading next. I have seen, firsthand, the value that perspective brings.”
Kurtis BartonSenior VP of Strategy & Sales, Voice Media Group
Approved for publication by Kurtis Barton, 2026-07-23.
“just wow...”
Kurtis BartonSenior VP of Strategy & Sales, Voice Media GroupUnprompted, on seeing a result. Email, 2026-07-18 21:41 UTC, subject line “wow”. This is the whole message — it is a reaction, not a testimonial, and it is not part of the quote above.
“And the hits just keep on coming!”
Jayson AkersDirector of SEO/AEO/GEO, V DigitalEmail, 2026-07-19 15:40 UTC.
The limits are part of the result.
This section is not a disclaimer. It is the part of the record that tells you how much weight the rest of it will carry.
No one can guarantee a specific answer to a specific prompt
Not us, not any vendor. What is engineered is retrieval, grounding and trust architecture — the conditions under which a model can find, parse and rely on a source. That raises the probability of a citation. It does not determine any individual generation.
Results vary by user, device, cache and RAG timing
On 2026-07-27 Kurtis Barton ran a prompt from a colleague's computer and got a different answer than he got from his own, and asked why. The honest answer is that the model frequently is not reading the live site — retrieval-augmented generation has its own refresh cadence, and even an explicit instruction to visit the page is not always honoured. Two people running the same prompt on the same day can legitimately see different results.
The headline compresses a sequence — it is not a controlled experiment
“Invisible to #1 in 3 days” is the observed order of events: omitted by five of five engines on the enterprise-SEO query at the 2026-06-17 baseline, then first on Perplexity and Grok at the 2026-07-16 audit, three days after go-live. The baseline asked two specific queries; the audit asked a third. A controlled re-run of the baseline queries has not been performed, which is why this page quotes no lift figure and no percentage anywhere.
Our own delivered artifacts are not evidence that the layer works
The machine-readable files now published on the client's domain — llms.txt, llms-full.txt, the MCP manifest, the entity record — are our work product. They belong in the description of what was shipped. They are not third-party validation and are not cited as such here. Every result on this page comes from an engine, a crawl log, or a probe.
Rankings move, and this one moved down
Six days after the audit that placed V Digital Services #1 on Perplexity and Grok and #2 on ChatGPT, ChatGPT returned it on neither of two prompts. That is on the sequence above with its date. A case study that showed only the upward days would be describing a different system than the one that exists.
Four measurements are out of band right now
SGR is 0.211 against a target of 0.25 or better. LMR is 53 days against a target of 30 or fewer. SEO sits at 57 and EGR at 33% — both softer than we consider settled. These are published beside the strong numbers because a panel with no weak readings is not a measurement panel.
Two figures are withheld outright
CITE reads 55% on the dashboard and 60 in the underlying store; we will not pick one. RPC displays 6,176 with no corresponding field in the store at all, so we cannot say what it counts. Both are named here rather than quietly dropped.
How to check any of this.
The baseline probe and its receipts exist as files — 30 of them across 10 screenshots, 10 raw API responses and 10 text transcripts, each raw response carrying a recorded SHA-256. The crawl log, the composite series and the instrumentation panel are read from the measurement store and dated at the point of reading.
Our measurement rules, and what we do when a number cannot be substantiated, are written down.