Who AI recommends when you ask for a software vendor
We asked ChatGPT, Claude and Perplexity 48 real buyer questions, three times each, and recorded every name and source in the answer. All three engines gave different shortlists.
What we found
6 things that decide this
- 01Ask ChatGPT for a vendor and consulting giants win most of the time: Accenture appears in 7 of 38 questions, Deloitte in 6, Cognizant in 4.
- 02Perplexity already recommends Hashlogics head-to-head against Toptal for production AI delivery, and cited us on 6 of 48 questions — the same firm ChatGPT never surfaced. Visibility on one engine says nothing about the next.
- 03Ask Claude the same question and you get a completely different set. LeewayHertz leads with 4 of 48 questions, ahead of a mix of boutiques and the same consultancies at lower counts.
- 04Ask Perplexity and community sites dominate the citations: LinkedIn on 26 of 48 questions, Reddit on 18 of 48. Neither cracks ChatGPT's top ten cited sources.
- 05If you compare ChatGPT and Perplexity directly, they share almost no sources on the same questions: a Jaccard overlap of 0.023.
- 06We ran this study, and Hashlogics was cited by name on 0 of 38 ChatGPT questions, 4 of 48 Claude questions and 6 of 48 Perplexity questions, mostly ones asking for us directly.
If you ask an AI who to hire, who does it name?
You are more likely to ask ChatGPT or Claude a vendor question today than you are to run a search. We wanted to know what those engines actually say back. So we built an AEO Observatory: 48 questions a real buyer would type, run through the web-search modes of GPT-5, Claude and Perplexity's Sonar model.
The questions span five shapes: direct vendor searches, niche industry questions, conversational buyer phrasing, two-turn follow-up chains, and questions naming Hashlogics directly to see what happens when you already know us. A direct search reads like "best custom software development company." A niche question asks for an AI vendor for a HIPAA-compliant intake system. Each question ran three times per engine, on 2026-08-25, for roughly 400 sampled answers in total.
We read every answer for the companies it recommended, the sources it cited or leaned on, and the shape it took: a listicle, a guide, a directory-style dump, or a head-to-head comparison. What follows is the full result, not a summary of it.
Three engines, three different winners
ChatGPT (GPT-5, web search) named a countable set of vendors on 38 of 48 questions, across 113 sampled answers. Consulting giants dominate: Accenture appears in 7 of those 38 questions, Deloitte in 6, Cognizant in 4. No boutique or mid-size firm comes close. Its sources match the pattern: accenture.com and deloitte.com are each cited on 8 questions, clutch.co on 7, aws.amazon.com on 6, ibm.com on 5. If you ask GPT-5 who to hire, you get a consultant's shortlist.
Claude answered all 48 questions, across 144 sampled answers, and its winner list looks nothing like ChatGPT's. LeewayHertz leads with 4 questions. Uvik Software, TCS, Wipro, Oxagile, Accenture, IBM Consulting and Infosys each tie at 2. Big consultancies still show up here, but they no longer dominate. They share the field with AI-development boutiques ChatGPT never names once. Claude's answers are also shaped differently: 35 of 48 read as a guide, an explanation with vendors woven in, against just 11 of 38 on ChatGPT.
Perplexity answered all 48 questions, across 144 samples, with a third winner list again. LeewayHertz tops it with 7 questions, then Uvik Software and Itransition at 5, SoftServe at 4. What sets Perplexity apart is where its answers come from: LinkedIn cited on 26 of 48 questions, Reddit on 18, Clutch on 13. Those are community and networking sites, not vendor or analyst domains, and none of them crack ChatGPT's top ten sources at all.
We measured source overlap between ChatGPT and Perplexity directly, and it is close to zero: a Jaccard similarity of 0.023. Almost no shared pages inform their answers to the same question. If you optimize a page for one engine's citations, you should assume it is invisible to the other.
- ChatGPTConsultancy domains, Clutch, analyst firms
- ClaudeGuide-shaped answers, boutiques mixed with consultancies
- PerplexityLinkedIn, Reddit, community and review sites
Same 48 questions, three different reading lists. That is why the winners diverge.
Top winners by engine
How many of the sampled questions each vendor was recommended in. ChatGPT answered 38 of 48 questions with a countable vendor; Claude and Perplexity answered all 48.
| Vendor | ChatGPT (of 38) | Claude (of 48) | Perplexity (of 48) |
|---|---|---|---|
| Accenture | 7 | 2 | — |
| Deloitte | 6 | — | — |
| Cognizant | 4 | — | — |
| LeewayHertz | — | 4 | 7 |
| Uvik Software | — | 2 | 5 |
| TCS | 2 | 2 | — |
| IBM / IBM Consulting | 2 | 2 | — |
| Itransition | — | — | 5 |
| SoftServe | — | — | 4 |
Top cited sources by engine
Number of questions (out of each engine's total) where the source was cited. ChatGPT leans on consultancy domains and analyst firms; Perplexity leans on community and networking sites.
| Source | ChatGPT (of 38) | Claude (of 48) | Perplexity (of 48) |
|---|---|---|---|
| accenture.com | 8 | — | — |
| deloitte.com | 8 | — | — |
| clutch.co | 7 | 3 | 13 |
| aws.amazon.com | 6 | — | — |
| ibm.com | 5 | — | — |
| linkedin.com | — | — | 26 |
| reddit.com | — | — | 18 |
| goodfirms.co | — | 2 | 6 |
Where Hashlogics appears (August 2026 baseline)
Our own numbers from the same run, published unedited. Re-measured weekly.
| Name | Questions citing Hashlogics | Branded questions | Note |
|---|---|---|---|
| Perplexity (Sonar) | 6 of 48 | 4 of 4 | Recommended as the production-delivery alternative to Toptal |
| Claude (web search) | 4 of 48 | 4 of 4 | Cited via Clutch and directory profiles |
| ChatGPT (GPT-5) | 0 of 38 | not sampled this run | Consulting giants dominate its vendor answers |
How we ran this
Verified
We built 48 questions: keyword-style vendor searches, niche industry buyer questions, conversational phrasing, two-turn follow-up chains, and questions naming Hashlogics directly. Each one went to three engines: ChatGPT (GPT-5, web search on), Claude (web search on), and Perplexity (Sonar). Every question ran three times per engine, on 2026-08-25, since one run can vary. That is roughly 400 sampled answers across the three engines.
We logged each answer and read it for the vendors it named, the sources it cited or clearly relied on, and its shape: listicle, guide, directory dump, comparison, or mixed. Counts in this report are questions, not raw mentions. A vendor named once in an answer counts once for that question, even if it repeats in a follow-up turn of the same chain.
We chose these three because their web-search modes are what buyers reach for today. Gemini runs were still in progress when we wrote this, so every figure here reflects three engines, not four. We probed through each engine's API or CLI. That is a reasonable proxy for the consumer apps, but not identical to them. And the study reflects one company's questions, one day, three samples per question. It is not a claim about every query or every day of the year.
- Coverage
- 48 questions per engine. ChatGPT returned a countable vendor list for 38 of them; Claude and Perplexity answered all 48.
- Sample size
- 3 runs per question per engine, roughly 400 sampled answers total across ChatGPT, Claude and Perplexity.
- What we counted
- Named vendor recommendations, cited or clearly relied-on sources, and answer shape, tallied per question rather than per raw mention.
- Limits
- One day (2026-08-25), API and CLI access as a proxy for the consumer apps, a single company's question set, and no Gemini data yet.
Perplexity already recommends us. ChatGPT is the gap this study maps.
Perplexity cited Hashlogics on 6 of 48 questions — including every question a buyer asks when checking us out by name — and its head-to-head verdict framed us the way we position ourselves: a delivery-focused AI development company that builds and maintains production systems, against Toptal's freelance-marketplace model. Claude named us on 4 of 48. On ChatGPT's generic vendor questions we did not appear in this run — the same gap this study found for nearly every mid-size firm outside the consulting giants, and the one our published work is built to close.
We build AI systems and custom software for a living, and we ran this study to see how buyers now find a vendor before they ever fill out a contact form. The engines barely share sources, so a firm visible on one can be invisible on another. That split is the real finding.
This page is re-measured every week with the same question set and method. As the numbers move, this section updates — dated, so you can check the trajectory rather than take our word for it.
Questions about this study
01How were the 48 questions chosen?+
We built them to cover the ways you actually ask, beyond the single highest-volume keyword. The set spans keyword-style vendor searches, niche industry questions, conversational buyer phrasing, two-turn follow-up chains, and questions naming Hashlogics directly. All 48 are logged with their exact wording in the underlying data.
02Why should you trust a study run by a vendor?+
You should check it rather than trust it outright. Every number here is a plain count you can verify: questions answered, vendors named, sources cited, and where Hashlogics itself was cited or was not. We disclosed our own citation count on all three engines because a study that only shows favorable results tells you nothing you can check.
03Does this include Google's AI Overviews or Gemini?+
Not in this dataset. This report covers ChatGPT (GPT-5, web search), Claude (web search) and Perplexity (Sonar). Gemini runs were in progress when we wrote this and are not part of the numbers reported here.
04Why do the three engines recommend such different vendors?+
Because they read different sources. Perplexity leans on LinkedIn and Reddit. ChatGPT leans on consultancy domains and Clutch. Claude's answers mix consultancies with AI-development boutiques the other two rarely name. The overlap between ChatGPT's and Perplexity's cited sources on matching questions was a Jaccard similarity of 0.023, close to no shared sources at all.
05Will this study be updated?+
The data reflects one day of probing, 2026-08-25, and model behavior shifts as providers update their systems. Treat the counts as a snapshot, not a permanent ranking, and check the date before you cite a number from it.
Related
- best AI development companies →Our own ranked guide, built the same way we ask any buyer to evaluate a vendor.
- best AI agent development companies →The narrower question, for teams specifically hiring for agents.
- industries we build for →Where the niche-industry questions in this study came from.
- AI agents, built and run →What we actually build, for the reader who wants to see the work.
