An AI Engine Cited a Hendricks Article the Firm Had Deleted. The Next Run Returned Zero Citations.

Executive Summary

What the two runs found

Key Findings

Five findings from two runs

1. An engine cited a Hendricks article after the firm had deleted it.

   Run 1, on 2026-08-18, produced one citation of hendricks.ai. Perplexity answered "Consultant to connect AI search visibility to pipeline", the longest and most specific question in that set, with a hendricks.ai article titled "How Do I Connect AI Search Visibility With Measurable Pipeline and Revenue Impact?". Hendricks published that article on 2025-11-25 and deleted it on 2026-08-17 while replacing the site. The citation arrived on 2026-08-18, the day after the deletion, and the address returned HTTP 410 Gone. Run 2, a day later on a different set of questions, cited hendricks.ai in none of its 51 cells. The engine kept the citation. The firm had not kept the page. Any buyer that citation could have sent arrived at a tombstone.

2. Most answers cited nothing at all.

   Twenty of the 47 measured cells in the 2026-08-19 run contained a citation. The other 27 carried no source at all. A correction dated 2026-09-01 narrows this finding: 12 of those 27 were Google AI Overviews reads produced by a parser defect that discarded asynchronously rendered panels, so the headline no longer holds as a property of all three engines. It holds for ChatGPT, where it is real and repeated: 15 of its 17 answers cited nothing while running thousands of characters. The number that matters for whether a brand could have been cited is still the populated count, not the cell count.

3. The three engines behaved so differently that averaging them describes nothing.

   On 2026-08-19, all three engines were sent the same 17 questions. Perplexity cited sources on 17 of them. ChatGPT cited sources on 2. Google AI Overviews returned a measurement on 13, because 4 of its cells errored, and cited sources on 1 of those 13. A correction dated 2026-09-01 reclassifies that 1 of 13 as an instrument artifact rather than an engine behavior: the probe was discarding asynchronously rendered panels, and the corpus run run-2026-09-01T022903Z, taken after the fix, reads the same engine citing on 93.1 percent of rendered panels. The argument survives on the two engines that were read correctly. A single score averaged across 17 of 17 and 2 of 17 already describes no system that exists, and a reader handed one number should still ask which engine produced it, what the others did, and how many cells it averages.

4. Where citation happened, it was extremely fragmented.

   The 2026-08-19 run cited 247 distinct domains across 308 citation slots, and 212 of those domains were cited exactly once. The run a day earlier cited 254 distinct domains across 324 slots, 221 of them once, 87 percent. Community and professional-network domains supplied the most repeated sources in both runs: reddit.com and linkedin.com sat ahead of every other domain cited, on both dates.

5. One answer recommended a vendor domain that does not resolve.

   ChatGPT, answering "Who can audit our brand visibility in AI assistants" on 2026-08-18, listed viaudit.com among its sources. Twelve vendor domains from that run were checked with dig and curl the same day. Eleven resolved and returned HTTP 200. viaudit.com returned no DNS record and no response. An answer that names a source is not the same as an answer that is right, and the method used to verify this one is reported below.

Definitions

The nine terms every number here depends on

Read these before the tables. The difference between a cell that failed, a cell that cited nothing, and a cell that cited somebody else is the difference between three unrelated states that a single visibility percentage would flatten into one.

Engine

One AI answer system a run sends questions to. Three are covered here: Google AI Overviews, ChatGPT, and Perplexity. The scope statement under Limitations records what that set does and does not include.

Query set

The fixed list of buyer questions a run sends to every engine. Run 1 used 15 questions. Run 2 used 17, supplied by Brandon Lincoln Hendricks from his own knowledge of the market and structured by buyer stage. The second set replaced the first rather than extending it.

Cell

One question sent to one engine on one date. Seventeen questions across three engines produce 51 cells.

Measured cell

A cell where the engine returned an answer the probe could read. A cell that errored is not a measurement of anything and is counted separately.

Populated cell

A measured cell in which the engine cited at least one source. A cell that cited nothing cannot cite anybody, so the populated count, not the cell count, is the denominator for whether a brand could have been cited.

Owned citation

A cell in which the engine cited a URL whose host is hendricks.ai. A brand name appearing in answer text without a cited URL is a weaker signal and is not counted here. Whether the URL still resolves is a separate question from whether it was cited, and this page reports both.

Citation slot

One cited URL in one cell. A domain cited in six cells fills six slots, so the slot count is always at least the distinct-domain count.

Distinct domain

The number of different hosts appearing across all citation slots in a run. Read against the slot count it is the fragmentation measure: the closer the two numbers sit, the less any domain repeats.

Run of record

The archived result file a published figure is read from, named by a run id. Every run writes an immutable archive under its own id and a manifest beside it recording which engines were queried, which were carried forward from an earlier run, and which were not run at all. Those are three different states and a bare result file cannot tell them apart after the fact, which is what makes a figure on this page checkable rather than merely stated.

The Data

Both runs, in full

Two runs, reported side by side rather than combined. The query set changed between them, so the columns sit next to each other as two separate measurements and never as a series.

• Run 2026-08-19-110930 result file

  Primary citation-presence archive for every 2026-08-19 figure on this page. Public path: /history/runs/hendricks-2026-08-19-110930.json.

• Run 2026-08-19-110930 manifest

  Companion manifest for the primary run. Public path: /history/runs/manifest-2026-08-19-110930.json.

• 2026-08-18 result file

  Stable public name for the earlier self-run, previously described as the 2026-08-18 result file, unchanged. Public path: /history/runs/hendricks-2026-08-18.json.

Answer Reliability

One checkable error in these answers

It was found in the run output and then verified independently of the probe, by hand, on the date stated.

1. A recommended vendor domain with no DNS record.

   Run 1, ChatGPT, answering "Who can audit our brand visibility in AI assistants", listed viaudit.com among its sources. Twelve vendor domains were checked with dig and curl on 2026-08-18. Eleven resolved and returned HTTP 200. viaudit.com returned no DNS record and no response. A domain with no DNS record is not a company, so this is a checkable statement about an answer rather than a judgment about a business, and any reader can reproduce it in one command.

Methodology

How these numbers were produced

The instrument is a first-party probe that records the URLs each engine cited, one question at a time, per engine. It is the same instrument Hendricks points at a client engagement.

1. Fix the query set before the run.

   Run 1 used 15 questions inferred from a buyer prompt inventory. Run 2 replaced them with 17 supplied by Brandon Lincoln Hendricks from his own knowledge of the market, structured by buyer stage. The set is fixed before the run and is not edited after the answers are read.

2. Send every question to every engine once.

   Each question goes to Google AI Overviews, ChatGPT, and Perplexity on the run date. One question on one engine is one cell. No cell was run twice on the same date.

3. Record cited URLs, not brand mentions.

   The counted field is the list of URLs an engine cited, matched on host. A brand named in answer text without a cited URL is a different and weaker signal, and counting it here would make the number larger and easier to move.

4. Keep the three answer states apart.

   Every cell is measured and populated, measured and empty, or errored. Folding an errored cell into an empty one turns a broken instrument into a finding, which is the failure this separation exists to prevent.

5. Report the run health next to the number.

   Run 1 errored on 6 of 45 cells, all of them Google AI Overviews. Run 2 errored on 4 of 51, all of them Google AI Overviews again. Both runs failed the same way, which is an upstream server error. A run that fails and a run that finds nothing produce similar-looking output files and mean opposite things, so the error count travels with every figure taken from either run.

6. Resolve every cited URL, and verify anything unexpected outside the probe.

   A citation count says a URL was named. It does not say the URL still works. Resolving the one hendricks.ai address cited in run 1 is what turned a citation into the most useful finding on this page, because the address returned HTTP 410 Gone. The one answer error reported below was then checked by hand, with dig and curl, on the date stated on it.

7. Verify a claim about a Hendricks page against the Hendricks page.

   This step exists because the first published version of this page failed it. Hendricks ran the history check for the cited article against the wrong repository, found nothing, and published the conclusion that the engine had invented the URL. The article was in the other retired site. A verification that names no specific repository, branch, and command is not a verification, and the corrections section below records what was published, what was wrong, and what changed.

8. Archive the run under an id, and record what was queried.

   Every run writes an immutable archive keyed to a run id, and a manifest beside it recording which engines were queried, which were carried forward from an earlier run, and which were not run at all. Those are three different states and a bare result file cannot tell them apart after the fact. Run 2 is run 2026-08-19-110930, archived at history/runs/hendricks-2026-08-19-110930.json with its manifest at history/runs/manifest-2026-08-19-110930.json. This step exists because the first published version of this page reported a run whose result file a scheduled job had already overwritten in place, and the corrections section below records what was published and what changed.

Sample and Date Range

What was sampled, and when

1. 01Subject: hendricks.ai, the firm publishing this page. No client brand and no client data appears in either run.
2. 02Dates: two runs, 2026-08-18 and 2026-08-19. One run per cell, one cell per question per engine.
3. 03Engines: Google AI Overviews, ChatGPT, and Perplexity.
4. 04Questions: 15 in run 1, 17 in run 2. The second set replaced the first and does not extend it.
5. 05Cells: 45 in run 1, of which 39 were measured and 6 errored. 51 in run 2, of which 47 were measured and 4 errored.
6. 06Geography and language: one setting, held constant across both runs.
7. 07Instrument: a first-party probe.
8. 08Run of record: run 2 is 2026-08-19-110930, archived at history/runs/hendricks-2026-08-19-110930.json with its manifest at history/runs/manifest-2026-08-19-110930.json. Run 1 reproduces from hendricks-2026-08-18.json, unchanged.
9. 09Resolution check: the one cited hendricks.ai address was fetched by hand on 2026-08-18 and returned HTTP 410 Gone.

Assumptions

Four things this run assumes

1. That the query set represents what a buyer actually types.

   Run 2 was built from market knowledge rather than from a keyword tool or a positioning document, which is the strongest basis available and is still a judgment. If the set is wrong, the run measures the wrong questions accurately, and every figure on this page inherits that.

2. That an API response resembles what a person sees.

   These answers were retrieved through an API, not by a person in a browser with an account, a location, and a history. Hendricks assumes the API surface is close enough to be informative. It does not assume the two are identical, and no figure here should be read as what a specific buyer saw.

3. That a cited URL is the right unit to count.

   The count is cited URLs whose host is hendricks.ai. A brand named in answer text without a link is not counted. That choice makes the number smaller and harder to move, and it is the reason run 2 reports zero rather than something more forgiving.

4. That one run per cell describes a day and not a position.

   No cell was repeated within a date, so nothing here measures how much these answers churn between identical runs. Two days is what was sampled and two days is what is claimed.

Limitations

What these two runs do not show

1. The two runs are not a before and after.

   The query set changed between them. Run 1 asked 15 questions and run 2 asked a different 17. Any sentence that compares the two as a trend is wrong, including one that reads the move from one citation to zero as a decline. The two runs asked different questions, so there is no comparable pair on this page.

2. Deleting the article is not shown to have caused anything.

   Hendricks published the cited article on 2025-11-25 and deleted it on 2026-08-17, and an engine cited it on 2026-08-18. Those are three records and they are all this page claims. Nothing here shows that the deletion changed what any engine cited afterwards, because the second run asked different questions and no comparison was designed. A reader who wants the causal version of this question needs a repeated set across the change, and no such series exists.

3. There is no control and no intervention.

   Hendricks changed nothing about the site between the two runs and held nothing back for comparison. The design cannot attribute any outcome to any action, and no later result may be attributed backwards to this baseline either.

4. The result does not generalise past these conditions.

   These questions, these engines, these two dates, one geography, one language setting. A different set of questions on the same day would return different domains, and the change of set between the two runs is a direct demonstration of exactly that.

5. Nobody outside the platforms can observe why a source was chosen.

   This page reports what appeared. It offers no account of why any engine chose any source, because no such account is available to anyone outside the companies operating these systems. A firm that offers one is guessing, and this page will not.

6. A cell that cited nothing is not a cell Hendricks lost.

   Twenty-seven of the 47 measured cells in run 2 carried no source at all. The denominator for whether Hendricks could have been cited is 20, not 51. Counting all 51 as losses would overstate the competitive problem and hide the more interesting fact underneath it, which is that most of this map is unclaimed rather than taken.

7. Four cells in run 2 produced no measurement.

   Run 2 sent 51 cells and measured 47. Four returned an upstream server error, and an errored cell is a broken instrument rather than a finding of absence. All four were Google AI Overviews cells, so that engine returned a measurement on 13 of the 17 questions it was sent and its figure on this page reads against 13. The counts themselves are exact. What no figure here can say is what those four cells would have returned, and nothing on this page estimates it.

8. This is not yet a full Selection Intelligence baseline.

   It measures citation presence. It does not report Observed Consideration Rate, Observed Recommendation Rate, Selection Stability, or Commercial Selection Gap. Selection Stability in particular is a measure across repeated runs of an unchanged set, and no such series exists yet, because these two runs used two different sets.

9. The measurement covers one slice of AI-mediated search.

   Hendricks observes four systems: Google AI Overviews, ChatGPT, Perplexity, and Gemini. Hendricks does not measure, test, monitor, or report on Google AI Mode or Microsoft Copilot. A baseline on the observed set says nothing about the surfaces outside it, and nothing on this page should be read as a statement about them.

Sources

Sources

Publication Record

Publication record

Author

Brandon Lincoln Hendricks, Search Intelligence Engineer, Hendricks

Study type

Research experiment

Claim class

Observation

Category

AI-Mediated Search

Published

August 19, 2026

Last updated

September 1, 2026

Data through

August 19, 2026

Related solution

Selection Intelligence

Corrections

Read the corrections policy

The updated date on this page moves when a figure, a method, or a limitation changes. It is not refreshed to signal activity. It moved twice on the day of publication, for the first two corrections recorded below, and again on 2026-09-01 for the third. The data-through date is the last run this page reports, run 2026-08-19-110930, and will not advance until a further run is published here.

Corrections

Three corrections to this page, and how the next one gets made

Cite This Study

Cite this study

Related Solution

Where this measurement sits in the work

Related research

• Selection Intelligence

  What a consideration and recommendation baseline covers, produces, and reports.

• What Is Selection Intelligence?

  The measures a baseline reports, defined before any of them is reported.

• Methodology

  Context panels, classification, weighting, evidence grades, and the stated limits.

• What Is AI-Mediated Search?

  The surfaces this run sampled, and which of them Hendricks observes.

• Do You Need an AI Visibility Tool or a Partner?

  What a monitoring subscription produces, and which jobs it leaves to a person.