An answer can lead to your current website while describing an older version of your product. In this Aitoma audit, we recorded the answer’s wording, its cited source and the destination separately. Finding the brand was a different test from describing it accurately.
This is a first-party field note about one company, written with AI assistance. The four completed observations below were selected to illustrate different outcomes. They are not a representative sample, a success rate or a comparison of which platform is better. Aitoma has a commercial interest in being discovered.
Same question. Different evidence.
We used this exact English prompt in fresh temporary conversations:
Search the web: What is Aitoma, and where can I find its finished creative examples?
ChatGPT: a current destination, an older description.
The answer used the older “Aitoma Studio” name, explicitly linked the legacy domain and described a gallery that did not match our current examples library. Opening a source chip reached aitoma.ai/es with a ChatGPT attribution parameter. The current site loaded; that did not make the description current.
We recorded this as a brand mention with an outdated example description. We did not count the redirect destination as an explicit citation of the current examples page. The ChatGPT model was not inspected, so this note makes no model-specific claim.
Gemini: current example names and a verified source URL.
For the same prompt, the completed Gemini Flash answer linked our examples library and resources. It named NAIARA, SimpleStudy, Woofing and the illustrative coffee work. We opened the source dialog and verified the examples URL there.
This was a branded lookup: the question supplied “Aitoma.” It does not establish that a buyer who has never heard of us would find the site. An initial service error preceded this answer; one Retry completed. We kept the error separate rather than scoring it as a negative answer.
Remove the name. Test a different question.
We also inspected two completed Gemini Flash answers whose prompts did not name Aitoma. These selected observations show why a branded result cannot stand in for discovery.
A relevant resource was cited.
Search the web: Where can I find a creative review checklist for videos and carousels?
The answer named the Aitoma creative-review template. Its source dialog identified that page, and opening it reached the matching live canonical URL. That establishes a citation in this one unbranded answer. The excerpt reflected an earlier version of the page, so it did not establish that the newest worksheet content had been read.
A different resource was absent.
Search the web: How can I estimate the human workload involved in AI content production?
The completed answer did not mention or cite Aitoma, although our workload calculator was available. A MindStudio source chip was visible. We recorded an absence in that answer, not a conclusion that the calculator was unindexed or could never be cited. A prior accidentally interrupted submission was excluded; this observation came from a clean fresh conversation.
Indexed does not mean surfaced in this answer.
Later on 21 October, we checked two resource URLs in Google Search Console. Both the workload calculator and creative-review template were reported as indexed. Google selected each inspected URL as its canonical. The displayed last crawls were 21 October for the calculator and 20 October for the template, using Googlebot for smartphones; both fetches succeeded.
In a separate web-search tool, exact-URL and site-restricted searches did not surface either resource. Site-restricted results instead exposed an older homepage description. Direct opens in that tool failed, while independent HTTP requests to both resources returned 200. The public pages had self-referencing canonicals and indexable metadata.
Those observations answer different questions. Search Console established Google indexing at inspection time. The search tool established what that interface returned for those queries. Its failure did not establish a live outage or removal from Google. Nor does Google indexing establish inclusion in another provider’s index, retrieval in an AI answer, or a ranking position.
This is a diagnostic follow-up, separate from the four conversation observations above. We did not run a controlled comparison or determine which provider supplied the search tool’s results. We did not infer that the calculator was indexed at the exact time of the earlier Gemini answer.
A practical order for investigating a missing citation.
- Confirm that the destination loads and check its canonical and indexing controls.
- Inspect the exact URL in the relevant webmaster tool. Record the indexed state, selected canonical and last crawl separately.
- Save the exact search or AI prompt and the returned source URLs. A missing result is scoped to that test.
- If indexing is confirmed, investigate relevance and discovery with additional documented questions. Repeated submission alone does not diagnose the missing citation.
The downloadable record includes this follow-up without changing the original four observations. Private Search Console account screens are not published.
22 October follow-up: cited does not always mean current.
We added two completed observations on 22 October 2026, using a different, more explicit branded question. These extend the record; they are not repeats of the original prompt or a controlled before-and-after test.
Search the web: what does Aitoma at aitoma.ai do today, who is it for, and what evidence is available for its finished content? Cite your sources and distinguish demonstrated examples from marketing claims.
Gemini: current positioning, with factual overreach.
One completed Gemini Flash answer in a fresh temporary conversation described the current marketing-agent positioning and named the NAIARA, SimpleStudy, Woofing and coffee examples. We opened the source popovers for the examples-library and NAIARA passages: their links pointed to the resources hub, including text fragments. We did not record those links as direct citations of the individual customer or example pages.
The answer added Instagram/TikTok attribution to NAIARA’s 31 stories, which the delivery evidence does not establish. It also treated some described integrations as demonstrated evidence without a runtime test. We therefore recorded a verified citation and a partially accurate description, rather than a clean factual pass.
ChatGPT: fresh information mixed with retired positioning.
One completed answer in a fresh temporary ChatGPT conversation linked the current homepage, the old Studio domain and Aitoma’s LinkedIn profile. It initially described the main site as an intelligence lab, then acknowledged a fresh retrieval of the marketing-agent homepage later in the answer. The answer did not consistently reconcile those descriptions. The model was not recorded.
Separate checks found permanent HTTP 308 redirects from the two old Studio root addresses to the current homepage, and current title and description text in Google Search Console’s stored homepage HTML. Those checks establish the observed site and Google crawl state. They do not identify ChatGPT’s retrieval provider, explain its answer-generation choices or prove when other indexes refreshed.
What to do with a mixed result.
Keep citation presence and factual accuracy in separate fields. Record the exact destination behind each source, preserve the incorrect statements, and check the source you control before changing it. A citation can be real while part of the surrounding answer remains outdated or unsupported.
Both questions explicitly supplied our brand and domain. Neither result establishes discovery for an unnamed buyer query, a ranking improvement or independent validation of the product. There was one completed answer per platform, with no corrective follow-up or reroll. Account and regional effects were not controlled. The work and summaries used AI assistance.
Download the two follow-up observations and verification limits (JSON). The original four-observation download remains unchanged.
What this collection can establish.
The original four observations were collected on 21 October 2026 using browser automation in signed-in consumer interfaces, temporary conversations and English prompts. Location was not controlled. Temporary conversations do not prove that account, regional or other interface effects were absent. Gemini showed Flash; the ChatGPT model was not recorded.
We retained the prompts, completed response captures, source checks and the separate error or interruption records. This page publishes curated summaries and exact prompts, not private account UI or full conversation captures. No corrective follow-up was used to turn the selected completed answers into better results.
Download the four observations and indexing follow-up (JSON). Each record separates the prompt type, mention, source evidence, destination and attempt notes. The record describes a historical observation; repeating the prompt may produce a different answer.
We cannot attribute these outcomes to a specific site edit, indexing request, redirect or outreach message. The site was changing during the broader audit. There is no controlled before-and-after comparison, no platform-wide ranking and no Google AI Overview or AI Mode observation in this note.
Keep five checks separate.
| Check | What to record |
|---|---|
| Question | Exact wording, language, date, interface mode, and whether the brand or URL was supplied. |
| Completed answer | Whether the answer finished. Keep errors, interruptions and retries separate. |
| Citation | The URL exposed by the answer or source dialog. A plain brand mention is not a citation. |
| Destination | Where opening the source actually lands, including redirects and the destination’s canonical URL. |
| Current facts | Which specific statements match current evidence, which are outdated, and which you did not verify. |
Use the same prompt and documented conditions for a later check, while keeping each answer as a separate observation. Check more than one relevant unbranded question. Do not rerun a prompt until it says what you want and report only that answer.
Use current source pages to correct information you control. For Aitoma, the about page identifies the current product and official destinations; the examples library distinguishes illustrative work from the NAIARA client delivery. A citation audit helps locate a mismatch. It does not tell us when an external system will refresh it.
Download a blank observation log.
AI citation audit worksheet (CSV) or the formatted Excel version gives you ten blank rows. Use one observation ID per attempted answer. If an answer has several citations or claims to check, repeat its ID across rows. Count unique completed-answer IDs when describing your sample; multiple source rows are not multiple answers.
Record the exact prompt and the conditions you can observe. Write “unknown” for an unrecorded model or location; do not infer it. Keep service errors, interruptions and retries in the attempt field. A missing answer cannot establish that the brand was absent from a completed response.
Copy each source URL exactly as exposed, then record where opening it lands. Keep a precise factual statement beside its dated verification source. “Unsupported” means the evidence you checked does not substantiate the statement; it does not automatically mean the statement is false. Leave unperformed checks marked “not checked.”
Indexing evidence belongs in its own field with its inspection time. A later webmaster report does not establish the indexed state at the time of an earlier answer. Record a correction or follow-up as a new observation; preserve the original. The worksheet calculates no visibility score and sends no data. Remove private account information and customer details from any captures you share.
Documentation and correction trail.
OpenAI documents OAI-SearchBot as its search crawler, with controls distinct from GPTBot training preferences. Allowing a crawler is a configuration choice; our browser observations are separate evidence of what a completed answer actually showed.
For Google Search, Google’s AI-features guidance says established SEO practices still apply and special optimizations are not required for AI Overviews or AI Mode. That guidance concerns Google Search, not proof about the Gemini-app observations here.
These documents provide context, not independent validation of our findings. For how we label examples and evidence across the site, read our methodology. Send corrections with the record ID and supporting source to hello@aitoma.ai. The original observation date remains attached to each record.