AI visibility audit

Executive summary

Nelix chip design appears in 20.9% of usable runs (Wilson 95% CI 16.7%–25.9%), based on 301 usable runs from 304 successful engine calls on 20 September 2026. Your own domain was cited in 57 runs, and the overall citation rate is 18.9%. Zero mentions is not the finding here; concentration is.

Two things matter commercially. First, all of your visibility sits in talent-pipeline and regional-identity questions: P04 (80.0%), P02 (76.5%), P15 (75.0%), P13 (64.7%), P14 (61.1%). Eleven prompts return 0% — including the ones that read like buying intent: P01 (RTL and verification outsourcing), P06 and P07 (hardware root of trust, secure boot consultants), P09 (cost of outsourcing SoC verification), P17 (choosing a design services vendor). Engines currently answer "who trains African chip engineers" with your name and answer "who should I hire" without it.

Second, where you do appear you lead: 53.85% share of voice against ChipMango Africa at 29.91%.

The rest of the report covers per-engine behaviour (Claude at 5.6% is the outlier), the competitive picture, the ranked source gaps engines lean on, and a phased 90-day work plan.

Engine by engine

Claude: 54 runs, 54 answers present (100% answer rate), 3 mentions — visibility 5.6% (CI 1.9%–15.1%), citation rate 1.9%. This is the sharpest asymmetry in the scan. Claude names you on P02 (Top fabless chip design startups in Africa) at 1.00 and nowhere else — it returns 0 on P04, P13, P14 and P15, all prompts you win comfortably elsewhere. Its share of voice puts ChipMango Africa (0.0741) ahead of you (0.0556).

Google AI Mode: 54 runs, 54 answers present, 12 mentions — visibility 22.2% (CI 13.2%–34.9%), citation rate 22.2%. You appear at 1.00 on P04 and P15 and at 0.6667 on P02, P13 and P14. Your share of voice on this engine (0.2222) is more than double ChipMango Africa's (0.0926).

Google AI Overviews: 34 runs, 31 answers present — a 91.2% answer presence rate, the only engine that did not return an answer on every call. 3 mentions, visibility 9.7% (CI 3.4%–24.9%), citation rate 9.7%. Here ChipMango Africa outranks you on share of voice (0.1613 vs 0.0968). You take 1.00 on P15 and 0.5 on P13, but 0 on P02.

Gemini: 54 runs, 54 answers present, 15 mentions — visibility 27.8% (CI 17.6%–40.9%), citation rate 27.8%. Joint strongest engine. Gemini returns you at 1.00 on P02, P04, P13, P14 and P15 — a clean sweep of the won prompts — and 0 on every commercial-intent prompt.

ChatGPT: 54 runs, 54 answers present, 15 mentions — visibility 27.8% (CI 17.6%–40.9%), citation rate 24.1%. The only engine that mentions you on P03 (FPGA prototyping and design validation as a service) at 0.3333. It also gives Synopsys University Program its highest share of voice anywhere (0.1111).

Perplexity: 54 runs, 54 answers present, 15 mentions — visibility 27.8% (CI 17.6%–40.9%), citation rate 24.1%. The only engine that mentions you on P11 (compute platforms designed for brownouts and intermittent power), and it does so on every run (1.00). No other engine mentions you on P11 at all.

The competitive landscape

Across the whole scan you hold 53.85% share of voice on 63 mentions (mention rate 20.93%). ChipMango Africa is second with 29.91% on 35 mentions (11.63%), then Synopsys University Program at 9.4% (11 mentions, 3.65%) and RISC-V International at 6.84% (8 mentions, 2.66%). Only four entities are tracked as named competitors in this scan; no others were measured.

You lead on four of six engines. On Google AI Mode your 0.2222 share beats ChipMango Africa's 0.0926; on Gemini 0.2778 vs 0.1111; on ChatGPT 0.2778 vs 0.1481; on Perplexity 0.2778 vs 0.1296. You are behind on two. On Claude, ChipMango Africa leads 0.0741 to your 0.0556. On Google AI Overviews, ChipMango Africa leads 0.1613 to your 0.0968.

The gaps split cleanly into two kinds. The Claude and Google AI Overviews deficits look incidental rather than structural: both engines already mention you, both sit on small mention counts (3 each) with wide confidence intervals, and on Google AI Overviews you actually take 1.00 on P15 and 0.5 on P13 — the shortfall is inconsistency, not absence. Claude's behaviour is the odder case: it recognises you as an African fabless startup (P02) but not as a Ghana design house (P04) or a curriculum partner (P15), which the other five engines do.

The structural gap is categorical, not competitive. On the eleven prompts you lose, the named competitor is Synopsys University Program on P01, P06 and P07 and nobody at all on P05, P08, P09, P10, P12, P16, P17 and P18. No rival owns those answers; engines simply have no vendor-level entity to attach to hardware root of trust, secure boot consulting, offshore RTL alternatives or SoC verification pricing that includes you. That is an evidence gap in the sources, not a share battle.

One observation worth flagging rather than interpreting: the scan geography is recorded as New Mexico, while every prompt you win is framed around Africa. The scan did not measure why that divergence exists.

The source-gap work plan

Ranked by citation frequency across all answers. The "co-cited with client" figure shows how often that domain appears in the same answer as you — a low figure on a frequently cited domain is where engines are building answers you are absent from.

• linkedin.com (cited in 15.3% of answers): the single most-cited domain, 46 runs, across Claude, Google AI Mode, Google AI Overviews, ChatGPT and Perplexity. Co-cited with you 54.4% of the time — so it is already carrying you about half the time it appears. A professional network profile is structured, frequently crawled and unambiguous about company identity, which is why engines reach for it. Action: audit and fully complete the company page — services taxonomy (RTL, verification, FPGA prototyping, secure boot), locations, and named specialisms matching the eleven lost prompts. Publish engineering posts under those terms.
• arxiv.org (cited in 7.6% of answers): 23 runs, five engines, but co-cited with you only 13.0%. A preprint server; engines lean on it for technical claims about edge inference, attestation and verifiable compute — exactly the topics of P18 and P05, where you score 0%. Action: publish or co-author technical preprints on verifiable edge inference and device attestation.
• fidus.com (cited in 7.0% of answers): 21 runs and the only domain cited by all six engines — and co-cited with you 0% of the time. A design-services vendor's own site. Engines treat it as a primary reference for what a design services engagement consists of. Action: you cannot claim this listing; instead build equivalent service-definition pages on nelix.io (scope, deliverables, timelines) so an engine answering P01 or P17 has a comparable source.
• anysilicon.com (cited in 6.6% of answers): 20 runs, five engines, co-cited with you 0%. An industry directory and reference resource for semiconductor supply and services — directories are cheap, comprehensive and structured, which is why engines cite them. Action: claim or request a company listing with full service categories.
• facebook.com (cited in 6.6% of answers): 20 runs, cited by Google AI Mode, Google AI Overviews and Gemini only; co-cited with you 40.0%. A general social platform. Action: ensure the page exists, is complete and consistent with LinkedIn and nelix.io; this is a consistency job, not a content programme.
• weforum.org (cited in 6.3% of answers): 19 runs, five engines, co-cited with you 47.4%. An institutional publisher; engines use it for macro framing on talent and industrial policy — the ground you already win. Action: pitch contributed commentary on semiconductor talent pipelines to sustain the co-citation rate.
• synopsys.com (cited in 6.0% of answers): 18 runs, five engines, co-cited with you 22.2%. Vendor documentation, and the home domain of Synopsys University Program, the competitor named on P01, P06 and P07. Action: no action available on the domain itself; treat it as the benchmark for the depth of technical documentation engines expect.
• youtube.com (cited in 5.7% of answers): 17 runs, cited by Google AI Mode, Google AI Overviews and Gemini; co-cited with you 17.6%. Video is heavily indexed by the Google-family engines. Action: publish short technical walkthroughs (FPGA lab setup, verification flow) with full descriptions and transcripts.
• einfochips.com (cited in 4.7% of answers): 14 runs, four engines, co-cited with you 7.1%. Another design-services vendor site. Action: as with fidus.com — build matching capability and case-study pages on nelix.io.
• tessolve.com (cited in 4.7% of answers): 14 runs across Claude, Gemini and ChatGPT, co-cited with you 7.1%. Design and test services vendor. Note that embedded.tessolve.com is separately cited in 3.0% of answers with 0% co-citation. Action: same mechanism — close the service-page gap on your own domain.
• startuplist.africa (cited in 4.3% of answers): 13 runs, five engines, co-cited with you 76.9% — the highest co-citation rate among the top gaps. A regional startup directory, structured and regularly refreshed. Action: verify and expand your entry; this source is already doing work for you.
• chipmango.com (cited in 4.0% of answers): 12 runs, five engines, co-cited with you 83.3%. A competitor's own domain; engines pull it into the same answers as you. Action: none directly — monitor only, as a signal that engines treat you and ChipMango Africa as a set.
• semidesignjobs.com (cited in 4.0% of answers): 12 runs, all six engines, co-cited with you 41.7%. A sector job board. Hiring listings are strong evidence of who actually does the work. Action: post live roles with detailed technical scope.
• utilitydive.com (cited in 4.0% of answers): 12 runs, five engines, co-cited with you 0%. A utility-sector trade publication — the source engines reach for on P08 (securing grid-edge devices) and P06 (smart meter attestation), both 0% for you. Action: pitch the editor a bylined piece on tamper resistance at remote grid sites.

The 90-day plan

Days 0-30

• [effort: low · impact: high] Complete and standardise the LinkedIn company page around the service terms in the lost prompts (RTL, verification, FPGA prototyping, secure boot, attestation) — evidence: linkedin.com is the most-cited domain in the scan at 15.3% of answers, 46 runs, across five of six engines, and is already co-cited with you 54.4% of the time.
• [effort: low · impact: medium] Claim and expand the anysilicon.com listing with full service categories — evidence: cited in 6.6% of answers across five engines with 0% co-citation with you today.
• [effort: low · impact: medium] Verify and expand the startuplist.africa entry — evidence: cited in 4.3% of answers with a 76.9% co-citation rate, the highest of any top-ranked source gap.
• [effort: low · impact: low] Align the facebook.com page with LinkedIn and nelix.io for name, location and service wording — evidence: cited in 6.6% of answers, and only by Google AI Mode, Google AI Overviews and Gemini.
• [effort: medium · impact: high] Publish service-definition pages on nelix.io for the eleven zero-mention prompts: RTL and verification outsourcing, hardware root of trust, secure boot consulting, SoC integration cost model, two-week architecture review — evidence: P01, P05, P06, P07, P08, P09, P10, P12, P16, P17 and P18 all return 0% mention frequency.

Days 31-60

• [effort: medium · impact: high] Pitch a bylined piece on grid-edge tamper resistance to utilitydive.com — evidence: cited in 4.0% of answers across five engines with 0% co-citation, and it is the trade source behind P06 and P08, both 0%.
• [effort: medium · impact: medium] Post detailed live engineering roles to semidesignjobs.com — evidence: cited in 4.0% of answers and the only job board cited by all six engines.
• [effort: medium · impact: medium] Publish technical walkthrough videos (FPGA lab build, verification flow) with full transcripts — evidence: youtube.com is cited in 5.7% of answers, exclusively by the three Google-family and Gemini surfaces where your Google AI Overviews visibility is weakest at 9.7%.
• [effort: medium · impact: medium] Build capability and case-study pages that match the structure of the vendor sites engines already treat as reference material — evidence: fidus.com (7.0%, all six engines, 0% co-cited), einfochips.com (4.7%, 7.1% co-cited) and tessolve.com (4.7%, 7.1% co-cited).
• [effort: low · impact: medium] Review Claude-facing public descriptions of the company for consistency with the P02 framing that already works — evidence: Claude sits at 5.6% visibility (CI 1.9%–15.1%) and mentions you only on P02, scoring 0 on P04, P13, P14 and P15, which every other engine wins.

Days 61-90

• [effort: high · impact: high] Publish or co-author a technical preprint on verifiable edge inference and device attestation — evidence: arxiv.org is cited in 7.6% of answers across five engines with only 13.0% co-citation, and P18 and P05 both return 0%.
• [effort: medium · impact: medium] Pitch contributed commentary on semiconductor talent pipelines to weforum.org — evidence: cited in 6.3% of answers with 47.4% co-citation, sitting on the topic ground where you already score 61.1%–80.0%.
• [effort: medium · impact: medium] Document a published engagement model and indicative commercial terms for outsourced SoC integration and verification — evidence: P09 and P17 return 0% and have no named competitor either, meaning no entity currently owns those answers.
• [effort: low · impact: medium] Extend the P11 material — the brownout and intermittent-power positioning — into written form on nelix.io — evidence: Perplexity mentions you on P11 in 100% of its runs while every other engine returns 0, suggesting the evidence exists in a form only one engine reaches.
• [effort: medium · impact: low] Build out the FPGA prototyping-as-a-service page specifically — evidence: P03 is contested at 5.9% overall, carried entirely by ChatGPT at 0.3333.

At day 90 we re-run the identical 18-prompt set against the same six engines at the same three runs per prompt per engine, so the recount is directly comparable to this baseline.

Methodology

The scan completed on 20 September 2026 and queried six live engine APIs: Claude, Google AI Mode, Google AI Overviews, Gemini, ChatGPT and Perplexity. All six returned status "ok". Eighteen prompts were run at three runs per prompt per engine, producing 304 successful runs of which 301 were usable for scoring.

Per-engine run counts and answer presence: Claude 54 runs, 54 answers present (100%); Google AI Mode 54 runs, 54 answers present (100%); Google AI Overviews 34 runs, 31 answers present (91.2%); Gemini 54 runs, 54 answers present (100%); ChatGPT 54 runs, 54 answers present (100%); Perplexity 54 runs, 54 answers present (100%). Google AI Overviews is the only engine with a lower run count and the only one that did not return an answer on every call — its figures therefore carry the widest confidence interval in the scan (9.7%, CI 3.4%–24.9%).

All confidence intervals are Wilson 95% intervals, which give an asymmetric range appropriate for proportions built on small counts — the correct reading of the headline 20.9% is "between 16.7% and 25.9%", not a point estimate.

Sentiment across mentions was classified as 31 neutral, 24 positive and 8 unclassified. Average first-mention position was 0.369. Your own domain was cited in 57 runs.

Caveat: these are API responses. Consumer apps for the same engines apply different retrieval, personalisation and ranking layers, so what a user sees in the app can differ from what the API returned here. Results are also a snapshot of one date.

Every raw answer from all 304 successful runs is stored verbatim and available on request, including the runs where you were not mentioned.