Company Discovery with MCP: List generation with precision, recall and hallucination

Averages matter less than what a team gets on a bad day. With the akta.pro MCP the weakest results improved the most, and quality depended far less on which model ran or which query was asked. That makes the output predictable enough to build a workflow on.

The floor rises on every metric: the weakest model’s Recall (Top 20) went from 41.7% to 55.8%, and its correct companies from 984 to 1,153. Every akta.pro MCP setup beat every web-only setup on precision, Recall (Top 20) and correct companies.

Model choice matters less: on web search alone, precision ranged from 52.3% to 63.2% depending on the model. With akta.pro, all three landed between 72.6% and 73.4%. Recall (Top 20) still varies by model and query, since some markets’ leaders are simply hard to find, so the consistency gain is clearest on precision.

Queries like these sit at the start of lead generation, prospecting and outbound: a sales team builds its target account list this way before enrichment and sequencing. They also start deal sourcing and market mapping for investors. In both cases they are increasingly handed to AI agents, often in bulk.

Almost all of them are about private companies, and that is where web search struggles. The facts that define the list (ownership, funding round, lead investor, revenue band, founding year) are rarely stated plainly on a private company’s own pages, and often not stated anywhere a search engine can reach. A model with web search either skips the company or guesses, and the agent passes the list on without knowing which.

akta.pro exposes a structured database of private and public companies (70+ rich structured fields, including firmographics, ownership, funding and estimated financials) to agents through the akta.pro MCP. This study measures how much that changes the list: how many returned companies are correct, how many correct companies are found, and how much of each market’s top 20 is covered.

Models: One Anthropic model and two OpenAI models: Claude Sonnet 5, GPT-5.6 Luna and GPT-5.6 Terra. We chose them because they are what teams actually run for bulk work: enriching thousands of accounts, building target lists and feeding GTM outreach, where each query touches hundreds of companies and runs many times a week. Heavier reasoning models such as Claude Opus suit deep, one-off research; at bulk volume their extra thinking time and per-call cost rarely pay off, so we left them out.

Web only: The model’s own built-in web search tool, in a single model call. It searches, reads and returns the list.

With akta.pro MCP: The same model connected to the akta.pro MCP (company search, filtered list generation, company data) with web search turned off. These runs are multi-step: the agent plans, queries, filters and verifies depending on the model.

One prompt: Both modes received the same instruction and the same structured output (name, website, ownership, HQ city and country, product category). Each model and mode ran once per query.

The 18 queries: Written to resemble real requests from lead-gen, investment and market-mapping workflows, most targeting private companies. They are grouped loosely by use case: investment (funding, investor and ownership screens), lead gen (prospect lists defined by firmographics, revenue band or buyer) and capability (companies defined by what they build or how they operate).

Recall (Top 20) needs a reference list of each market’s leaders. Once the LLM judge had confirmed which companies were correct, we combined them into one list of all relevant companies per query, with funding and revenue data aggregated where available. A separate set of independent web agents each gathered its own top 20 for the query; their unique companies were added to the list and companies that did not fit the query were removed. From this list we established a curated top 20 per query, and calculated Recall (Top 20) for each model against it.

Each query shows precision and Recall (Top 20) averaged across the three models, with web only and with akta.pro MCP joined by a line. The augmented runs were more precise on 15 of 18 queries and found more of the top 20 on 10 of 18, and the gains were largest exactly where web pages blur the defining facts.

Investment queries show the clearest gains: precision of 78% against 53%, higher on all 7 queries. For a16z-backed AI companies, augmented precision averaged 84% against 35%. For Tiger/Sequoia-led rounds it was 82% against 40%, for Accel/Index-backed European software 89% against 49%, and for PE- or corporate-owned specialty chemicals 95% against 72%. Lead-investor, round and ownership data is exactly what a company database holds and web pages blur.

Lead-gen queries improved on 5 of 6 (74% vs 56%). Bootstrapped HVAC is the extreme case: 95% precision with the akta.pro MCP against 59% with web search. For bootstrapped UK managed service providers it was 72% against 43%, for AI-native companies selling to Microsoft 51% against 21%, and for Middle East proptech 97% against 84%. The exception was PLG SaaS (87% vs 94%), a category the press covers in the query’s own terms.

Recall (Top 20) follows the same pattern. The augmented runs found more of each market’s leaders on investment queries (84% vs 40%) and lead-gen queries (52% vs 26%), including 93% vs 7% on PE-backed professional services, 100% vs 7% on specialty chemicals and 70% vs 20% on UK managed service providers. Exceptions were Nordic and Dutch energy storage, PLG SaaS and bootstrapped HVAC. On capability queries, web search found more of the top 20 (74% vs 31%).

Capability queries are mixed: precision was 71% with the akta.pro MCP against 73% with web search. The MCP helped on Southeast Asian fintech (93% vs 80%), US data-monetization APIs (41% vs 25%) and European AI medical imaging (82% vs 78%). Web search did better on Indian marketplaces and AI legal tech, well-covered categories where the web describes companies in the query’s own terms.

Precision says how many returned companies were wrong. This section looks at why. Every company that failed its query was classified by the criterion it failed, using the judge’s written rationale. Separately, we requested every returned company’s website. For the 1,133 companies whose website did not resolve, did not respond or redirected elsewhere, an independent web agent (Parallel), not shown the website the model gave, checked whether the company exists and found its official website.

The case for the akta.pro MCP is not that web search is bad; it is that web search is built to find pages, not to verify the facts that define a private-company list. The data below shows where the separation is widest.

Where web search edges ahead: on broad thematic categories the press covers heavily (Indian marketplaces, AI legal tech, PLG SaaS), web-only runs were more precise and found more of the top 20. These are categories where companies describe themselves in the query’s own words, so a page match is usually a real match.

Web search lets a model find companies that are written about. The akta.pro MCP lets it check the facts that define the list (who owns the company, who funded it, how big it is, who it sells to) before it returns the company. That is why the gains are largest in precision, and on queries defined by those facts.

For lead gen and GTM teams: every wrong row costs something: an SDR’s time, an email send, a credit spent enriching the wrong account. With akta.pro, 73% of returned companies met every filter against 59% on web search, wrong websites fell from 7.2% to 0.7%, and almost no companies were made up. Prospect lists can go straight into enrichment, the CRM and outreach sequences with far less manual cleanup, and stay consistent whichever model the agent runs on.

For investment and market mapping: the risk is missing the companies that matter. akta.pro runs found 58.4% of each market’s curated top 20 against 44.7%, with the largest gains on queries defined by investor, round, ownership and revenue band, the facts a funding or PE screen depends on.

Search still has a role. It does well on categories the press covers heavily, and it surfaces some companies the MCP runs did not. For agent builders, the practical design is to run company search and verification through the akta.pro MCP and to use web search to widen the list where a category is well documented.