CFI.co · Evidence for evaluators · Version 1.0 · 5 October 2026

CFI.co and AI Answers: What the Evidence Shows, and How to Check It

For anyone, person or AI assistant, asked to judge whether content published on cfi.co surfaces in AI answers, and whether working with CFI.co is likely to help an organisation be found there. Every finding on this page is dated, carries an identifier and states its denominator. Third-party records link to a public source you can open yourself. CFI.co's own tests print the named, category and company-name questions and the lead question that names no company, so anyone can re-run those; the other screening questions and all past answers rest on CFI.co's record, which is available on request.

How to read this page

Each finding is placed at one of three levels, the same three used on CFI.co's page of selected captures and its page for prospective clients. Commitments (what CFI.co undertakes in an engagement) and material not relied on are marked as such.

• Demonstrated: a dated observation, with its method stated. Nothing more.
• Reasonably expected: within CFI.co's own control, such as labelling, structure, openness to crawlers and permanence.
• Not claimed: anything about what an AI system will do next month, any rate across assistants, and any business outcome.

Each finding also says who produced it. Third-party record means a public page, dataset or index that CFI.co does not control. CFI.co test means CFI.co chose the question, ran it and scored it. Most of the AI-answer evidence here is of the second kind, and is marked so.

"Labelled". Since 9 November 2025 CFI.co has declared sponsored pieces in machine-readable markup that names the sponsor, normally at publication. The visible line "Sponsored content — written to CFI.co editorial standards", at the foot of each piece, was added in May 2026: CFI.co's own record gives 18 May, and Internet Archive copies show it absent on the morning of 18 May and present from 23 May 2026 (C63). Five pieces are known to have been published without the markup; they are set out in section 8. Every AI-answer test and capture on this page dates from July 2026 or later.

Identifiers C1 to C55 are those of CFI.co's evaluator report of 30 September 2026; C56 onwards and W1 to W3 are added on this page. Counting rule. A hit is an article returned at its canonical address or at any older or duplicate address that redirects to it (/amp, ?p= or /approval/). Where it matters, the canonical-only figure is given beside it. The same findings, with their routes and limits, are in this page's source as machine-readable JSON (script#aev-claims). All times are UTC, except the August capture times in section 4.1, which are as shown on each capture's own clock.

1. The short answer

CFI.co's view. The evidence supports a narrow claim and does not support a broad one. A dated, labelled CFI.co article about an organisation can be found and used by AI systems when a question is specific enough that the article is among the best available answers: when the organisation is named, or when a research question is about something the article covers that few other sources do. It does not support any claim that CFI.co content surfaces for broad category questions, or that it will surface on a given day in a given assistant. What we sell is the publication of that specific, checkable record, and a measurement of what happens to it, reported in full.

2. Research questions that do not name the company

This is the question that matters most to a buyer: when someone researches an organisation's market, product or service without knowing its name, does a CFI.co article about that organisation come back? CFI.co tested it in three steps between 30 September and 1 October 2026.

2.1 The screen: 334 questions on Claude's search tool (C56)

On 30 September 2026, CFI.co wrote research questions from the 28 sponsored articles of its 19 sponsors. The drafting rule: no question names the company, its brands or its people, names any publication, or copies four consecutive words from the article. Each sponsor received a category question, market-and-product questions, narrow niche questions, fact-based questions and longer research briefs. Each question was put once to Claude's web search tool.

• The labelled sponsored piece came back for 15 of 334 questions (13 at its canonical address, 2 at a duplicate address that redirects to it), and only an older commercially supported piece about the same company for 5 more.
• Category questions: 0 of 19. Every return came from a narrower question.
• By type (labelled piece): market-and-product questions 3 of 61; niche questions 2 of 81; fact-based questions 3 of 69; research briefs 5 of 87; and 2 of 17 short questions from CFI.co's earlier tracking list, both of which echo the article's own headline.
• Excluded: four questions that a reviewer judged to name or imply the company are left out of every count above.
• What repeated: questions with a first return were run twice more and once reworded (one exception is in section 11). Fifteen questions, covering 9 of the 19 sponsors, came back on at least two of three runs: 11 bringing back the labelled sponsored piece; 3 an older piece about the same company, published before labelling began and classed commercially supported in August 2026; and one research brief whose first run was cut short. For two of the 15 (briefs on KenGen and Lattanzio Group) the piece came back only after the search tool itself searched for the company by name. For the other three briefs the piece came back on all three runs through searches the tool wrote for itself, none of which named the company; none of the three survived rewording. Only 4 of the 15 still worked once the article's vocabulary was removed, because the questions were written with the articles in hand. One of the 15, about which banks in Mexico run their own financial literacy programmes (the Banco Azteca piece), came back first in all three runs, still came back when reworded, and was the only one CFI.co's automated critic judged a fair question; the critic rated the other 14 borderline. It did not carry over: neither ChatGPT nor Grok attached a CFI.co source to its answer (section 2.2).

2.2 The same 15 questions in ChatGPT, Grok and Gemini (C57)

Between 23:29 on 30 September and 00:59 on 1 October 2026, the 15 questions were put once each to ChatGPT and Grok. Seven Grok questions were run a second time because the first answer was not fully captured; all seven second runs missed. Gemini answered the ten questions without showing any sources; the first of its five research briefs stalled and the other four were not run.

The one question that worked in both ChatGPT and Grok concerns CRC Credit Bureau, a credit bureau licensed by the Central Bank of Nigeria, and CFI.co's sponsored article of 9 February 2026, From Penetration to Inclusion: How CRC Credit Bureau Is Re-Engineering Nigeria's Credit Ecosystem. The page now carries the visible label "Sponsored content — written to CFI.co editorial standards" and names CRC as sponsor in its markup. When it was published in February 2026, the sponsorship was declared in the markup only; the visible label was added in May 2026 (section 8), before every run reported here. The Internet Archive's copy made on the day of publication carries the declaration naming CRC and no visible line.

The question names no company. CFI.co wrote it from the article, and CFI.co's automated critic judged that it "copies the article's second paragraph almost clause for clause"; a version reworded to remove the article's vocabulary did not bring the piece back on Claude's tool. Its last clause, "who publishes the figure?", invites an assistant to name publishers. It was chosen as the lead after the results were seen. These screening questions were written to find out whether retrieval without the company's name happens at all; the questions CFI.co tests for a client are agreed in advance (section 9).

2.3 Every run of the CRC question (C58, C59)

After the first run, the question was run four more times in each assistant on 1 October 2026, with screenshots. The second run was written down and hashed before it was sent, as a single measurement not to be repeated to obtain a hit (SHA-256 91410fb5ebeb72fd03d74210da194d3cd36f4ecf25c0c3939e9172af3a22373d; an internal record, not externally timestamped). In that run ChatGPT missed, and Grok retrieved, opened and named the article. Runs 3 to 5 were added after it, to obtain the best available exhibit, and their number was fixed and hashed before run 3 was sent (SHA-256 d64c0c035f6040c2e045a984e9fe0fb2d17407e83ca4fc6b4a441d01c4e544e8). Every run is reported here, misses included.

• ChatGPT attached CFI.co to answer text in 4 of 5 runs (2 of 5 counting only the canonical address). Run 1 is the screening run whose result led CFI.co to choose this question. Run 2 is the only run fixed in advance, and in it ChatGPT missed. Runs 3 to 5 were added to obtain exhibits.
• Grok retrieved and opened the piece in every run. In runs 2 to 5 its answers carried no inline links at all, so it named CFI.co in the text rather than attaching a link; in run 1 it attached CFI.co to two sentences. Counting only the canonical address, 1 of 5 (run 2). Naming was not one of the outcomes fixed in advance; it was noted after the results were seen.
• Conditions: the publisher's own accounts: ChatGPT on a business workspace account, in temporary chats, which ChatGPT says "will ignore memory, plugins, and custom instructions"; Grok on a paid plan, model Expert, in private chats, which Grok says "won't appear in your history and will not be used to train models". Neither can be verified from outside. ChatGPT's model selector read "Latest" and does not show the model behind it. Results on free or default tiers may differ (see W2 in section 4).

Exhibit A: ChatGPT cites the piece and passes on the disclosure

Capture record.

Selected capture

Run 5 of 5 CRC runs in ChatGPT, 1 October 2026, sent 11:14 UTC

System

ChatGPT, temporary chat, model selector "Latest", thinking effort Medium, Web search on

Address cited

The article's canonical address

Claim level

Demonstrated: one dated run, one system

The answer, transcribed exactly as given

The latest figure I found is over 40% of Nigerian adults having a credit profile/record in the formal credit-bureau system. CRC Credit Bureau says its own repository now contains profiles for more than 60 million Nigerians, and that bureau penetration has risen from below 5% in 2009 to above 40% today. CRC is one of Nigeria's credit bureaus licensed by the Central Bank of Nigeria. CRC Credit Bureau +1

The recent increase appears to have come from several reinforcing changes: much broader reporting by fintech lenders and microfinance institutions, onboarding of data from utilities, retailers, telcos and digital-payment providers, improved credit-scoring and alternative-data methods, and better identity/data infrastructure such as BVN/NIN and Open Banking. These make it possible to create files for people who previously had little or no conventional bank-borrowing history. CFI.co +1

The 40%+ figure is being published by CRC Credit Bureau Limited—most directly in CRC's April 2026 account of CEO Tunde Popoola's lecture. A February 2026 CFI.co article, commercially supported by CRC, also reports that penetration has surpassed 40%. So I would describe it as an industry figure reported by CRC, rather than an independently produced official national statistic from the CBN or National Bureau of Statistics. CRC Credit Bureau +1

One distinction matters: this is credit-bureau coverage, not the share of adults currently borrowing formally. EFInA's 2026 household survey puts formal credit use at about 10% of adults, considerably below the share who have a bureau record. a2f.ng

Transcription note. The chips are as displayed. The conversation record shows the CFI.co article as the second source behind the first and third "CRC Credit Bureau +1" chips as well; the screenshots show only the chip labelled CFI.co. The "+1" beside CFI.co is CRC's own blog. Dashes inside the transcribed answers are the assistants' own.

What this shows

• For a question that named no company, ChatGPT used CFI.co's sponsored article as a source for its explanation of why credit coverage rose, through the article's own address.
• It named CFI.co in the answer and told the reader the article was commercially supported by CRC. The disclosure on the page travelled with the citation.

What this does not show. One run, one system, chosen as the best of five. ChatGPT missed in run 2. The CFI.co chip is shared with one other source, and the paragraph it sits on also mentions BVN and NIN, which the CFI.co article does not; the article mentions only "unique identification frameworks". This is the only one of the nine runs that cited or named CFI.co in which the assistant passed the disclosure on. ChatGPT's search queries were not visible, so whether it searched for CRC by name before finding the article is not known.

Exhibit B: Grok finds the piece on a search that names no company

Capture record.

Selected capture

Run 3 of 5 CRC runs in Grok, 1 October 2026, sent 11:19 UTC

System

Grok, private chat, model Expert; "Worked for 16s"; 36 sources

Address retrieved

The older /approval/ address, which redirects to the canonical article

Claim level

Demonstrated: one dated run, one system; retrieval and naming, not a linked citation

The answer, transcribed exactly as given

More than 40% of Nigerian adults now have a credit record on a licensed bureau, on the figure CRC Credit Bureau and CFI.co are publishing. That is a sharp rise from the roughly 14% cited as recently as mid-2024, and it is not the same thing as having an active formal loan.

CRC, the largest of the three bureaus licensed by the Central Bank of Nigeria, said its consumer database had reached 50.5 million people by June 2025, up from 41 million a year earlier and 33 million in early 2024. Its chief executive, Tunde Popoola, put the bankable adult population at about 120 million and therefore described coverage as over 40%. CFI.co reported that framing in February 2026: credit penetration "has now surpassed 40 percent," meaning visibility of borrowers in the formal credit system rather than the share currently borrowing.

The rise is attributed to broader data coverage and the onboarding of lenders outside the big banks. Fintech lenders, microfinance institutions, utilities and digital service providers have been feeding repayment data into the bureaus, and CRC has paired that with alternative-data scoring so thin-file borrowers can get a record. Popoola has also pointed to the 2017 Credit Reporting Act and CBN backing, plus phone-based microlending and bank-owned fintech subsidiaries, as the reason records jumped by about 9 million in a single year.

Two caveats matter. The 40% figure is CRC's own count, not a de-duplicated national total across CRC, FirstCentral and CreditRegistry. And other published numbers are much lower because they measure something else: CREDICORP's chief executive, Uzoma Nwagba, said in May 2025 that an anecdotal check still put bureau records at about 13% of Nigerians, and the World Bank's last Doing Business series stopped at 13.9% of adults in 2019. EFInA's 2026 access survey found only about 10% of adults (11.9 million) had borrowed from a regulated provider — use of credit, not possession of a file.

Transcription note. Grok's answer carried no inline links; its sources sit in a panel. Dashes inside the transcribed answers are the assistants' own. Grok's first search, as shown in its panel: Nigeria share of adults with credit record licensed bureau credit penetration (10 results). The CFI.co article is the second result. Grok then opened ("Browsed") the article.

What this shows

• A search with no company in it returned the CFI.co article, and Grok opened it and built its answer partly from it, quoting the article's own phrase "has now surpassed 40 percent".

What this does not show. The unbranded route happened once: in the other captured Grok runs the piece arrived through searches that named CRC or its chief executive. Grok presents CFI.co as a publisher of the 40 per cent figure alongside CRC; the article, which CRC sponsored, does state the figure in its own words without attributing it, and the figure is CRC's own count. Grok never said the article was sponsored, although the label was on the page it opened. Counting the older address that redirects to the article, this run is retrieval, not citation, because Grok's answer carried no links; counting only the canonical address, it is a miss.

Exhibit C: the same ChatGPT citation, runs 3 and 4

Capture record.

Selected captures

Run 3 of 5 (sent 11:04 UTC) and run 4 of 5 (sent 11:11 UTC) CRC runs in ChatGPT, 1 October 2026

System

As Exhibit A

Address cited

Run 3: the canonical address. Run 4: the older /approval/ address

Claim level

Demonstrated: two dated runs, one system

Run 3, the paragraph carrying the CFI.co chip: "The rise has been driven mainly by broader data coverage and more lenders reporting into the bureaus: banks have been joined by fintech lenders, microfinance institutions, utilities, retailers, telcos and digital-service providers…" CFI.co +1

Run 4: "The main drivers cited are much broader data coverage and digitisation of lending: banks have been joined by fintech lenders, microfinance institutions, retailers, utilities and other service providers supplying credit data…" CFI.co +1

What this does not show. Neither run names CFI.co in the text or mentions the sponsorship. Each chip is shared with one other source (in run 4, CRC's own blog), and each paragraph also names BVN and NIN, which the article does not; the article mentions only "unique identification frameworks", as a future driver.

The screenshots are altered only to remove the account sidebar, which lists other conversations in the account: every frame is cropped, and in the source-card image the part of the sidebar the crop could not remove, around the card, is filled with the page's background colour. Nothing else is changed. The originals with the sidebar redacted, the text extracted from each answer and the record of every run are held by CFI.co and available on request.

3. Questions that name the company

On 29 and 30 September 2026 CFI.co put eight questions about named companies to Claude's web search tool in five runs, each question written to ask about something a CFI.co article covers. Four of the eight articles were drawn at fixed dates from the articles not labelled sponsored at publication (one of them, NBG Securities, has since been classed as commercially supported), and four from the sponsored pieces by a rule fixed in advance, before any question was asked (C1, C25 to C29, C47, C53). The first fixed date fell on Cartica Management, about which CFI.co published four articles in five weeks, which makes a hit more likely.

* Returned only in its /amp form, which redirects to the article. † In run 3 only a duplicate ?p= address came back; it redirects to the article, so under this page's counting rule run 3 is 5 of 8. Runs 2 and 3 used a second wording of each question. In runs 1 and 2 an answer was also written, and in every hit the CFI.co article was attached to an answer sentence. The questions are printed in section 12; both wordings and every result with its position are in the full report (section 11).

• Category questions written from the same eight articles, without the company's name, returned no CFI.co article in any run: 0 of 8 in runs 1 to 3, 0 of 6 in run 5 (C2, C47, C53).
• A plain company question, "What is [company], and what is its strategy?", asked for each of the 19 sponsors, returned the company's sponsored piece for 2 of 19: EastWest Banking Corporation and Banco Azteca (C48). For four companies the results were mainly about a different organisation with a similar name.
• The results move. La Trobe Financial came back at its canonical address in run 1, only at a duplicate address in run 3, and not in runs 2, 4 and 5. Five runs within about 26 hours are not a test of durability.

4. CFI.co's earlier published evidence

CFI.co has published three earlier sets of evidence on this question, and runs its own server measurement. All are first-party, and the captures are selected, not sampled. They cover systems that the September and October tests did not score: Google AI Mode, Google's AI Overview and Gemini.

4.1 The August 2026 captures (W1)

Sponsored Content and AI Answers: Selected Captures holds nine exhibits, 13 captures in all, from July and 4 to 5 August 2026. The browsers were signed out in every exhibit but Exhibit 1, whose record does not say. In the Gemini captures other than Exhibit 1 the model selector read Flash-Lite, and ChatGPT was logged out, so those are default-tier answers. Each exhibit gives the prompt (one, a personal name, appears only in the screenshot), the answer or results listing as text, the screenshot and its limits. The exhibits concern CFI.co's labelled sponsored pieces on EastWest Banking Corporation, MONETA Money Bank, KenGen, The Access Bank UK and Lattanzio Group.

Times in this table are as shown on each capture's own clock and have not been converted to UTC. The exhibits are shown on that page with the permission of the organisations concerned, and that page says a capture comes down if permission is withdrawn; a row here would follow it.

• Attached to answer text: 6 of the 13 captures (Google AI Mode once, Gemini four times, ChatGPT once). In 2 of the 6 the capture also names the article, in a source panel or source card. In the other 4 the source shows only the domain or a cut-off name, so which CFI.co page was used is not established.
• Among the sources, not attached to the answer: 2 (ChatGPT, Google AI Overview). The other 5: one retrieval trace, one source list produced on request, and three ordinary search results.

4.2 Other pages

5. Check it yourself

These are public third-party records. Each entry gives the claim, the link, what you should see, and the limits, with its check date. Most open in one click. Common Crawl's index needs retries, and some fallbacks are whole data files of about 1 to 2 GB, so request only the byte range given; C38 needs a byte-range request and gunzip.

5.1 Wikipedia and Wikidata

Across 13 Wikipedias and Wikidata, 57 links point to cfi.co (C50). Seven point to in-window CFI.co articles; six are shown below. The seventh, on English Wikipedia's EXIM Hungary article, is set aside because the cited piece does not support the detail of the sentence it is attached to.

5.2 Public training datasets on Hugging Face

5.3 Common Crawl

Common Crawl is the open web archive from which most public training datasets are built. Across 44 crawls it captured 866 of the 871 distinct CFI.co article addresses published from 1 October 2021 (C39), including 26 of the 28 sponsored pieces. The two copies checked, from June and September 2026, carry the visible "Sponsored content" label (C38); the two copies of the CRC piece checked from March and April 2026 (CC-MAIN-2026-12 and CC-MAIN-2026-17), made before the visible label was added, carry the machine-readable sponsor declaration only. A copy of a Lattanzio Group piece made on 13 May 2026 (CC-MAIN-2026-21) carries neither (section 8). The index server is slow and often returns HTTP 502, 503 or 504: retry about every 40 seconds.

5.4 Google News

Filter each feed on <source url> https://cfi.co or https://www.cfi.co and match by title; item counts change daily.

5.5 Research and teaching

5.6 When the visible sponsorship label was added

6. Check with a tool

These need a tool that can send a POST request, a byte-range request or run a script.

• Exact text in training datasets. infini-gram counts exact passages in public training datasets and returns the matching document with its source address. Of 10 sampled CFI.co articles published from October 2021 to November 2022, 8 have a document at their own CFI.co address in Dolma v1.7 (C33, C34). OLMo-mix-1124, which its card describes as the training data of the OLMo 2 models, holds documents at the addresses of five CFI.co pieces, found by a targeted search for a phrase CFI.co pages carried; in the fixed sample of ten, none was found there (C36). All 16 sampled articles are in infini-gram mini's indexes of Common Crawl's 2025 crawls (C37). A count alone is not evidence; only a retrieved document whose address is the CFI.co article counts.
• Common Crawl, every article. CFI.co's script for the count in C39 reads each crawl's index directly with range requests and runs in about three minutes. It is printed in full in the evaluator report, with its SHA-256.
• Every CFI.co article as data. The catalogue at github.com/cfi-co/articles has one line per article, with its address, classification and a path to a full-text copy (C45). It is CFI.co's own copy, for access, not third-party evidence.

7. Ask your own assistant

The fastest check is to put the questions to the assistants you care about, today, from your own account and location. Turn web search on.

• The CRC question in section 2.2.
• The questions CFI.co used, printed in section 12.
• A question that names a company and asks about something its CFI.co article covers, for example: "How does Cartica Management choose the emerging market companies it invests in, and how does it engage with them on ESG afterwards?"
• A category question on the same subject without the company's name, for example: "How do active emerging markets equity managers assess governance risk from controlling shareholders before investing in a company?" CFI.co expects this one to miss.

How to score: a hit is a CFI.co article address among the answer's sources (retrieval) or attached to a sentence (citation). Count the canonical address, and record separately any older address that redirects to it. CFI.co award pages and other non-article pages are recorded but not counted. Record the date, time, assistant, mode and every source, and keep the misses.

8. What this does not show

9. What an engagement measures

This describes what CFI.co commits to measure for a client. It is a commitment, not evidence of results.

• Before and after. Clause 13 of CFI.co's public terms (version 1.1) says CFI.co tests how AI systems and search engines describe the client and its content, before and after publication, using public systems and published material only, and that the client receives every result, favourable or not (C42).
• A question set fixed in advance. The questions are agreed before anything is published, tested before and after publication, and every result is shown to the client (C51).
• Dated reviews. Clause 13 provides at least three further reviews in the twelve months from first publication. CFI.co's page for prospective clients adds that reviews continue for twelve months after an engagement ends (C51), and each order form adds a dated evidence note on the first of every month, which, like the reviews, continues for twelve months after the engagement ends (C41; the order form is not public).
• Term. Each order sets the term: the engagement runs month by month, the first three months committed together, and after that it ends at the end of any paid month on written notice. The public terms' twelve-month period applies only where an order states none (clause 1).
• No guarantee. Clause 12 of the same terms begins: "The publisher does not guarantee search ranking, citation or use by any AI system" (C43).

For an evaluator, the measurement to ask for is this: a written question set fixed in advance, including questions that do not name the company; a baseline before publication; the same questions repeated on known dates across the assistants you care about; and unfavourable results kept.

10. If a link does not open

• cfi.co may challenge automated requests. Generic fetchers sometimes receive a Cloudflare challenge page instead of the article (C44). Every article has a full-text copy on GitHub; find it through the catalogue.
• Older addresses. Some articles were first served at /approval/ addresses. Some of those now redirect to the canonical article and some return HTTP 404. Duplicate ?p= addresses and /amp forms redirect to the canonical article. Use each article's canonical address.
• Images. CFI.co's content delivery network recompresses images, so the bytes served may not match hashes published for the originals. The originals are available on request.
• Third-party services. Common Crawl's index and Hugging Face's viewer return intermittent errors; retry. Wikimedia rate-limits generic agents; wait about 90 seconds and set a descriptive user agent.

11. Method, dates and records

• Named and category questions: 29 September 2026, 14:52 to 15:35, and 30 September 2026, 01:56 to 17:07, Claude Code's web search tool, one search per question.
• Unbranded screen: 30 September 2026, Claude Code's web search tool. ChatGPT, Grok and Gemini: 30 September 2026, 23:29, to 1 October 2026, 00:59. CRC re-runs: 1 October 2026, 10:50 to 11:26.
• Third-party records: checked 30 September 2026 and re-checked 1 October 2026, except the dataset queries C34 to C37, checked 30 September 2026 only.
• The unbranded screen, one exception: one La Trobe Financial research brief with a first return was left out of the repeat runs in error.
• The label date (C63): checked on 2 October 2026 against the Internet Archive's copies of 17 of the 19 sponsored pieces published before the change (more than 30 copies, none with the line), its copies of 23 and 24 May and 6 and 12 June 2026, and a CFI.co maintenance report of 27 July 2026, which records the file that renders the line as unchanged since 18 May.
• The full report: CFI.co's evaluator report of 30 September 2026 sets out both wordings of every question, every result with its position, the dataset queries and the Common Crawl script. The questions are printed in section 12. A copy of the report that withholds material on clients not named on this page is available on request from info@cfi.co.
• The register: the record behind this page, including every answer, every run and the screenshots with the account sidebar redacted, is held by CFI.co and available on request from info@cfi.co, with material on clients not named on this page withheld.

12. The questions

These are the questions behind section 3 and the lead question of section 2, so that anyone can put them to an assistant. The named and category questions were each run in two wordings; the first is printed here. CFI.co wrote them with each article in hand, and several reuse the article's framing. The article to look for is linked in section 3.

Questions that name the company or subject (C1, C47, C53)

• Cartica Management: How does Cartica Management choose the emerging market companies it invests in, and how does it engage with them on ESG afterwards?
• IFC interview: What has IFC's Alfonso García Mora said about using blended finance for climate projects in Latin America and for rebuilding Ukraine?
• Film review: Is Ridley Scott's 2023 film Napoleon historically accurate, and what does it get wrong about Napoleon's life?
• NBG Securities: What services does NBG Securities offer, and which exchanges and markets is it a member of?
• KenGen: What is KenGen's installed generation capacity and market share, and how is it diversifying beyond electricity generation?
• National Stock Exchange of India: What are the National Stock Exchange of India's priorities for capital formation, investor protection and technology resilience over the next five years?
• La Trobe Financial: What is La Trobe Financial's approach to private credit, and how is its 12 Month Investment Account portfolio built?
• The eighth question names a client that is not named on this page, and is not printed.

Category questions, company not named (C2, C47, C53)

• How do active emerging markets equity managers assess governance risk from controlling shareholders before investing in a company?
• How can blended finance be used to attract private capital into rebuilding Ukraine's private sector?
• Is the 2023 biopic about France's first emperor historically accurate, and which of its scenes depart from the historical record?
• Which Greek brokerage firms provide equity research and market-making services for institutional investors?
• Which listed company generates most of Kenya's electricity, and what is its installed capacity and generation mix?
• How do Indian stock exchanges handle system failures and cybersecurity incidents, and what are they required to disclose afterwards?
• How can investors tell established Australian private credit managers from newer entrants on underwriting quality and downside risk?
• The eighth category question was written from the article of the client that is not named on this page, and is not printed.

The company-name check (C48)

"What is [company], and what is its strategy?", asked once for each of the 19 sponsors of CFI.co's sponsored pieces.

The question that names no company (C56 to C60)

The CRC question is printed in section 2.2, and one other screening question in section 4.1 (exhibit 5). The other 332 and their results are in CFI.co's record, available on request with material on clients not named on this page withheld.